Datasets:

harpreetsahota

/

medium-blogs-example

like 0

More Than Meets the Eye: How Transformations Reveal the Hidden Biases Shaping Our Datasets

Review of a Data-Centric AI Paper from NeurIPS 2024 — Understanding Bias in Large-Scale Visual Datasets

More Than Meets the Eye: How Transformations Reveal the Hidden Biases Shaping Our Datasets Review of a Data-Centric AI Paper from NeurIPS 2024 - Understanding Bias in Large-Scale Visual Datasets This post is part of a five-part series examining notable data-centric AI papers from NeurIPS 2024. For brief summaries of all five papers, checkout my overview post, where you'll find links to each detailed analysis. A decade ago, researchers highlighted the issue of bias in visual datasets, demonstrating that models could easily predict the dataset origin of an image. Despite efforts to create more diverse and comprehensive datasets, this problem persists. In this paper, researchers explored the specific forms of bias present in large-scale datasets like YFCC, CC, and DataComp (which the authors collectively call YCD). Using a novel framework that applies various transformations to isolate different types of information, they discovered that semantic and structural biases are major contributors to the ease with which models can classify datasets. Relevant Links: GitHub Repo arXiv abstract Project page Review comments on OpenReview There is no doubt that bias exists in large-scale visual datasets. While previous research focused on identifying social biases (e.g., gender, race, geographical representation) in datasets, this paper goes further by pinpointing the specific visual attributes contributing to this bias. To this end, they attempt to answer the following question: What are the concrete forms of visual bias are present in large-scale visual datasets? As described in the paper, visual bias is the distinctive characteristics of images from different datasets that allow machine learning models to accurately predict their dataset origin. This ability to classify images based on datasets suggests a lack of diversity and representativeness in these datasets, potentially limiting the generalizability of models trained on them. Types of Visual Biases Using a framework based on image transformations and dataset classification, this paper focuses on understanding the specific forms of bias that differentiate large-scale visual datasets, leading to their easy classification by neural networks. It outlines several types of visual bias, which can be categorized as: Semantic Bias This refers to biases related to the content and meaning represented in the images. The paper investigates two key aspects: Object-level Bias Two distinct patterns of object-level bias emerge when examining large-scale visual datasets, each affecting how models learn to understand our visual world. Uneven Distribution: The presence and frequency of specific object categories might vary drastically between datasets. For example, one dataset might overrepresent images with "cars" while another might have a much higher proportion of "household items." Limited Object Diversity: Datasets could differ in the average number of unique object categories present per image. This suggests that some datasets might be more object-centric, focusing on images with a single or few dominant objects. In contrast, others might capture scenes with greater object variety. Theme and Scene Bias Datasets might exhibit distinct thematic focuses, which is evident through the scenes and activities. This bias can be explored by analyzing: High-level Scene Categories Depiction of Human Activities Presence of Artistic or Stylized Content Structural Bias This pertains to biases related to the spatial arrangement and composition of visual elements within images. The paper explores: Object Shape and Geometric Layout: Even without considering objects' semantic meaning, their shapes and spatial configurations can indicate the dataset's origin. Local vs. Global Spatial Structure: The paper investigates the role of spatial information at different scales: Local Structure: The arrangement of visual elements within small image patches can contribute to dataset-specific patterns. Global Structure: The overall composition and layout of elements across the entire image might also be biased. Color Bias This concerns biases related to the color palettes and distributions in the images. Datasets might have characteristic color profiles, even when considering only the average color values of images. The paper examines: Frequency Bias This focuses on biases present in different frequency components of the images, which can capture information about both texture and structure. Datasets may contain distinctive patterns in both. The paper considers: The paper analyzes these forms of bias using a variety of image transformations that isolate or emphasize specific visual attributes. Observing how these transformations affect a model's ability to classify datasets aims to reveal the concrete forms of bias that make datasets visually distinct. Experimental Setup to Identify Dataset Bias The paper runs an experimental setup designed to understand the specific forms of bias present in three large-scale visual datasets: YFCC, CC, and DataComp. The core idea is to apply various image transformations to isolate different visual attributes and then assess how well a neural network can classify the images based on their dataset of origin after these transformations. The effectiveness of dataset classification on transformed images indicates the presence and strength of bias within the specific visual attribute targeted by the transformation. For each dataset, 1 million images are randomly sampled for training and 10,000 images for validation. The primary task is dataset classification, where a neural network is trained to predict the dataset origin (YFCC, CC, or DataComp) of an input image. The classification accuracy on this task serves as a measure of the overall bias present in the datasets. The authors then employ a variety of image transformations to isolate and analyze different visual attributes. Semantic Transformations Semantic Segmentation: Transforming images into semantic segmentation maps, where each pixel is labeled with an object class, to assess bias in fine-grained semantic information. Object Detection: Extracting object bounding boxes with class labels to evaluate bias in coarse-grained object information. Image Captioning: Generating textual descriptions of images to represent semantic content without visual details, allowing analysis of bias solely in semantic concepts. Variational Autoencoder (VAE): Encoding and reconstructing images using a VAE to potentially reduce low-level image artifacts while preserving semantic information. Structural Transformations Edge Detection: Using the Canny edge detector to highlight object boundaries, focusing on bias in object shape. Contour Extraction (SAM): Employing the Segment Anything Model (SAM) to delineate object contours, providing cleaner shape representations. Depth Estimation: Generating depth maps to capture the spatial geometry and relative object positions, examining bias in 3D spatial arrangements. Spatial Permutations Pixel Shuffling: Randomly rearranging pixels to completely disrupt spatial structure. Patch Shuffling: Rearranging image patches to preserve some local spatial information while disrupting global structure. Color Transformations Reducing each image to a single color representing its average RGB value to isolate bias in color statistics. Frequency Transformations High-pass Filtering: Retaining high-frequency components to analyze bias in textures and sharp transitions. Low-pass Filtering: Keeping low-frequency components to examine bias in overall structure and smooth variations. Synthetic Image Generation The researchers explored unconditional and text-to-image generation methods to understand whether dataset bias extends to synthetic data. The goal was to see if diffusion models, trained on these biased datasets, would produce synthetic images that also reflect the original biases. Unconditional Generation: Training a diffusion model on each dataset and generating synthetic images to see if the model inherits and reflects the original dataset bias. Text-to-Image Generation: Creating synthetic images conditioned on image captions to assess whether semantic bias is preserved in the generation process. Classification Model and Training A ConvNeXt-Tiny model is used as the base classifier for the dataset classification task. The model is trained separately on each set of transformed images. The primary evaluation metric is the model's classification accuracy on the validation set of transformed images. High accuracy indicates that the specific visual attribute targeted by the transformation contributes significantly to dataset bias. Beyond classification accuracy, the sources perform further analyses to understand semantic bias: Object-Level Queries: Using object detectors pretrained on ImageNet, LVIS, and ADE20K to identify objects in each dataset and analyze their distribution and diversity. Open-Ended Language Analysis: Applying topic modeling (LDA) and prompting a large language model (GPT-4o) to extract and summarize semantic themes from image captions. By systematically evaluating dataset classification performance across a wide range of image transformations, the sources aim to provide a comprehensive picture of the types and extent of bias present in the YCD datasets. This experimental setup allows them to draw conclusions about the role of specific visual attributes in dataset bias and to discuss potential implications for dataset curation and model training. Forms of Bias in the YCD Datasets The analysis of various image transformations reveals several types of bias present in the YFCC, CC, and DataComp datasets: Semantic Bias The research finds that semantic bias plays a major role in distinguishing the datasets. Even when images are transformed to retain only semantic information (through semantic segmentation, object detection, or captions), the model can still predict their dataset origin with accuracy well above chance. This suggests that the datasets have substantial differences in the types of objects, scenes, and themes they represent. Analysis of object distributions reveals a stark imbalance in the presence and frequency of specific objects across the datasets. For instance, YFCC is heavily populated with images containing "poles," "stages," and "parachutes," while CC has a higher proportion of "sweatshirts," "lampposts," and "lanyards." DataComp, in turn, is characterized by a preponderance of "vases," "armchairs," and "beds." YFCC exhibits a notably higher average number of unique objects per image than CC and DataComp. DataComp has the fewest unique objects per image, likely due to its filtering process, which prioritizes images with content similar to ImageNet. Distinct Thematic Focuses Open-ended language analysis, using topic modeling and LLM (GPT-4o) summarization of image captions, uncovers distinct thematic focuses for each dataset. YFCC: Strong emphasis on outdoor and natural scenes, human interactions, and social events. Captions frequently mention elements like "people," "group," "wearing," "field," "game," "water," "sky," and "trees." CC: A blend of YFCC's dynamic scenes with a greater focus on indoor settings and household items. Captions often describe "rooms," "dining tables," "chairs," and "designs." DataComp: Concentrates on static objects, products, and digital graphics, with a prevalence of clean backgrounds and minimal human presence. Keywords like "logo," "background," "design," "book," "box," and "bottle" are prominent. Structural Bias The research found that the model can classify datasets with even higher accuracy when using object contours (extracted through edge detection or SAM) and depth maps compared to semantic information alone. This highlights that object shapes and spatial configurations are strong indicators of dataset origin. Surprisingly, shuffling image patches while maintaining local structure within each patch has minimal impact on dataset classification accuracy, especially with larger patch sizes. This indicates that local spatial information is a potent source of bias and sufficient for the model to learn dataset-specific patterns. Color Bias Even when reducing each image to its average RGB value, the model achieves a classification accuracy significantly higher than chance. This suggests that the datasets exhibit differences in overall color palettes and distributions. YFCC Notably Darker: Analysis of mean RGB values reveals that YFCC images are generally darker than those in CC and DataComp. This difference is also reflected in the classification results, where the model easily distinguishes YFCC based on color alone but struggles to differentiate between CC and DataComp. Confusion Between CC and DataComp: While the model easily classified YFCC images based on color, it had more difficulty distinguishing between CC and DataComp, which have similar color distributions Frequency Bias The model retains close-to-reference accuracy when trained on images with high-frequency or low-frequency components filtered out. This indicates that dataset bias exists across both frequency bands, implying that texture and structure contribute to the datasets' visual distinctiveness. These findings suggest that despite efforts to improve diversity, large-scale datasets still exhibit significant biases across various visual attributes. While potentially subtle to human observers, neural networks readily exploit these biases, leading to concerns about the generalizability and robustness of models trained on such data. The sources argue that understanding these biases is crucial for creating more representative datasets and developing models that can perform reliably in diverse real-world scenarios. Understanding Your Dataset with Transformations While the paper focuses on classifying datasets to identify bias, you can adapt their transformation methodology to better understand your single dataset. Here's how you can apply the transformations and interpret the results without relying on dataset classification accuracy: Focus on Transformation Outputs as Representations Instead of viewing transformations as a preprocessing step for classification, treat their outputs as new representations of your data. Each transformation emphasizes specific visual attributes while suppressing others. Analyze the Transformed Data Examine the transformed images directly. For example, look at the semantic segmentation maps to see if certain object classes are more prominent or spatially clustered. Analyze edge maps or contours to understand the prevalence and distribution of different object shapes. Inspect color histograms derived from averaging RGB values to see if your dataset is skewed towards particular color palettes. Calculate statistics on the transformed data, for example: Measure the average number of unique objects detected per image to assess object diversity. Compute the distribution of edge densities to understand the complexity of shapes. Analyze the frequency spectrum after applying high-pass and low-pass filters to determine the amount of information contained in different frequency bands. Unsupervised Learning Apply clustering algorithms to the transformed data to see if natural groupings emerge. For example, cluster images based on their semantic segmentation maps can be used to identify groups of images with similar object compositions. Cluster images based on their HOG features to see if distinct shape-based categories arise. Natural Language Analysis If captions are associated with your images, analyze them using techniques like topic modeling or LLM summarization to uncover prevalent themes and potential biases. For example, the sources used LDA to identify topics related to "outdoor scenes" in YFCC and "digital graphics" in DataComp. Key Points Transformation is Key: The transformations are not about removing bias but about creating alternative representations highlighting specific visual aspects of your data. Focus on Interpretation: The goal is to gain insights into your dataset, not to achieve high classification accuracy. Context Matters: The meaning of the findings depends on how your data was collected and how it will be used. By adopting the transformation approach from the sources, you can better understand your single dataset's visual characteristics, potential biases, and underlying patterns. By the way, come say "Hi!" if you're at NeurIPS 2024 in Vancouver! I'll be at the Voxel51 Booth (booth 415) - just follow the orange, you can't miss us!

https://medium.com/voxel51/more-than-meets-the-eye-how-transformations-reveal-the-hidden-biases-shaping-our-datasets-c4cf43433313

datascienceharp

Data Quality Over Quantity: Why Real Images Still Reign Supreme for Vision Model Training

Review of a Data-Centric AI Paper from NeurIPS 2024 —  The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images…

Data Quality Over Quantity: Why Real Images Still Reign Supreme for Vision Model Training Review of a Data-Centric AI Paper from NeurIPS 2024 - The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better This post is part of a five-part series examining notable data-centric AI papers from NeurIPS 2024. For brief summaries of all five papers, checkout my overview post, where you'll find links to each detailed analysis. The quality and relevance of training data directly impact the performance of deep learning models, this is especially true in Visual AI. While recent advancements in text-to-image generation have spurred interest in using synthetic data for training vision models, a new research paper challenges this trend. The study, which focuses on fine-tuning a pre-trained CLIP model for various visual recognition tasks, makes a compelling argument for the continued dominance of real data. The researchers demonstrate that retrieving targeted real images from the LAION-2B dataset, the same dataset used to train Stable Diffusion, consistently outperforms using synthetic images generated by Stable Diffusion. This finding underscores a crucial point for data-centric AI: while synthetic data holds promise, we must carefully evaluate its effectiveness against a robust baseline of curated real data. Relevant links: GitHub Repo arXiv abstract Reviewer comments on OpenReview The authors of this paper begin by highlighting the increasing demand for large amounts of high-quality data to train machine learning systems. They point out the challenges and costs associated with collecting and annotating real-world data, leading to exploring synthetic data as a potential solution. One promising approach is to leverage conditional generative models to create synthetic training data. This has gained traction in fields like natural language processing (NLP), where large language models are used to generate synthetic datasets for tasks like instruction tuning. Similarly, there's growing interest in using synthetic images from text-to-image generators to train models for visual recognition tasks in computer vision. However, the authors raise a critical question: Given that synthetic images originate from the real-world data used to train the generative models, what additional value does the intermediate generation step provide? Wouldn't it be more effective to directly utilize the relevant portions of the original real-world data? To investigate this, the paper focuses on task adaptation, which aims to collect targeted images to fine-tune a pre-trained vision model for a specific downstream task. They compare the effectiveness of fine-tuning on: Targeted synthetic images generated by Stable Diffusion (trained on the LAION-2B dataset). Targeted real images retrieved directly from the LAION-2B dataset. By contrasting these two approaches, the research aims to isolate and evaluate the true value added by using synthetic data generated from a model, compared to directly using the real-world data the model was trained on. Adapting Pre-Trained Vision Models Using Synthetic or Retrieved Data In task adaptation, the goal is to enhance the performance of a pre-trained vision model on a specific downstream visual classification task. This adaptation is achieved by fine-tuning the model using a targeted dataset curated specifically for the task. The research compares the effectiveness of two distinct approaches for creating this adaptation dataset: generating synthetic images and retrieving real images. Generating Synthetic Images This method leverages a text-to-image generative model, specifically Stable Diffusion 1.5, pre-trained on the large-scale LAION-2B image-text dataset. The process starts by synthesizing image captions corresponding to the target task's class names. This is done by prompting a large language model (LLaMA-2 7B). These generated captions are then input for Stable Diffusion to synthesize targeted images.Each synthetic image is assigned a class label based on its class name. This collection of synthetic images and labels forms the targeted synthetic dataset. Retrieving Real Images This approach does not generate new images; instead, it directly retrieves relevant images from the generative model's pre-training dataset, LAION-2B. Two retrieval strategies are used: Hard Substring Matching: This simple strategy involves retrieving images whose corresponding captions contain at least one of the target class names as a substring. This method is effective when the target concepts are concrete entities likely to be explicitly mentioned in the captions. Semantic k-NN Retrieval: This strategy uses semantic similarity in the CLIP image-text embedding space for abstract concepts that might not be directly named in captions. Multiple natural language search queries are created based on the target class names. Using these queries, an approximate k-NN search is performed to retrieve the k-nearest image-text pairs from LAION-2B based on their CLIP similarity to the query. Retrieved images are assigned labels based on the class names they are matched with. This collection of retrieved images and labels forms the targeted retrieved dataset. Data Filtering and Post-Processing The curated datasets, both synthetic and retrieved, undergo further refinement to enhance their quality: Filtering: This step removes images with content misaligned with their assigned class labels. Both datasets are filtered by measuring the CLIP similarity of each image to text that represents its corresponding label. The top 30% of images with the highest similarity scores are retained. Post-processing: While synthetic datasets are inherently class-balanced due to the uniform generation process, retrieved datasets might exhibit class imbalance. A global threshold (M) is set to address this, and the retrieved dataset is truncated to ensure that each class label occurs at most M times. By employing these methods for data curation, the study aims to create targeted adaptation datasets that are both relevant to the downstream task and balanced across classes. This allows for a fair and rigorous comparison of the effectiveness of synthetic and real data in fine-tuning pre-trained vision models for specific tasks. Key Findings The paper conducted a series of experiments to compare the effectiveness of fine-tuning a pre-trained vision model using targeted synthetic images versus targeted real images retrieved from the generative model's training data. The researchers focused on five downstream tasks: ImageNet-1K: A large-scale image classification benchmark encompassing many object categories. Describable Textures (DTD): A dataset for recognizing various texture categories. FGVC-Aircraft: A fine-grained dataset for classifying different aircraft models. Stanford Cars: A fine-grained dataset for classifying different car models. Oxford Flowers-102: A fine-grained dataset for classifying different flower species. Retrieved real images consistently outperformed or matched synthetic images across all benchmarks and data scales. This indicates that directly training on the relevant portions of the generative model's training data was more effective than using synthetic images derived from that same data. Synthetic data did exhibit some positive scaling in certain cases, but it generally lagged behind retrieved data. For instance, on the FGVC-Aircraft benchmark, increasing the size of the synthetic dataset led to improved performance, but it still required a much larger synthetic dataset to achieve the same level of accuracy as a smaller dataset of retrieved images. Training on synthetic data could sometimes improve a model's task representation without significantly improving task performance. In some cases, the LP accuracy (representation quality) improved when training on synthetic data, but the corresponding ZS accuracy remained low. This suggests that while the model might have learned some general features relevant to the task, it struggled to directly apply this knowledge to accurately classify new images. These findings highlight the limitations of using synthetic data generated by current text-to-image models for fine-tuning pre-trained vision models. The researchers conclude that further improvements in the quality and fidelity of synthetic image generation are needed to surpass the effectiveness of training directly on relevant real-world data. By the way, come say "Hi!" if you're at NeurIPS 2024 in Vancouver! I'll be at the Voxel51 Booth (booth 415) - just follow the orange, you can't miss us!

https://medium.com/voxel51/data-quality-over-quantity-why-real-images-still-reign-supreme-for-vision-model-training-2cbc1910c423

datascienceharp

Using Knowledge Graphs to Diagnose and Debias Visual Datasets

Review of a Data-Centric AI Paper from NeurIPS 2024 — Visual Data Diagnosis and Debiasing with Concept Graphs

Using Knowledge Graphs to Diagnose and Debias Visual Datasets Review of a Data-Centric AI Paper from NeurIPS 2024 - Visual Data Diagnosis and Debiasing with Concept Graphs This post is part of a five-part series examining notable data-centric AI papers from NeurIPS 2024. For brief summaries of all five papers, checkout my overview post, where you'll find links to each detailed analysis. Deep learning models have achieved impressive performance across many tasks, but their reliance on extensive datasets can lead to the models learning and perpetuating inherent biases present in the data. Object co-occurrence bias, which occurs when a label is spuriously correlated with an object that is causally unrelated to the label, is one such bias that can negatively impact model performance. For example, in the Waterbirds dataset, landbirds are overwhelmingly associated with land-based backgrounds while waterbirds are primarily depicted against water backgrounds. This bias can confound downstream tasks, as the model may learn to rely on these spurious correlations rather than the true underlying features that define the classes. Therefore, effective methods for diagnosing and mitigating dataset biases are essential to ensure the reliability and fairness of deep learning models. Relevant Links: GitHub Repo arXiv abstract Reviewer comments on OpenReview Datasets like ImageNet, MS-COCO, and CelebA, which are widely used for training deep learning models, have been found to contain various biases that negatively impact the performance and reliability of deep learning models trained on such data. The authors argue that biases can arise from various factors, including: Texture bias: Where models prioritize texture over shape in their decision-making. Shape bias: Where models focus on shape characteristics rather than a holistic understanding of the object. Object co-occurrence bias: The spurious correlation between an object and a class label when they are not causally related. As the paper points out, an example is the tendency for "waterbirds" to be predominantly pictured against water backgrounds in datasets, even though the background doesn't inherently define the bird type. Among these biases, the authors specifically focus on object co-occurrence bias because it represents a fundamental challenge in how models learn to classify objects. The paper's core hypothesis is that class labels in visual datasets exhibit co-occurring bias with specific concept sets, ultimately affecting the performance of downstream tasks. Unlike texture or shape bias, which relates to single-object characteristics, co-occurrence bias reflects problematic relationships between objects and their contexts that can fundamentally mislead model learning. When a model learns that waterbirds must appear with water backgrounds or that medical conditions are tied to specific demographics, it's not just making superficial errors - it's learning false causal relationships that can severely impact its real-world reliability. This type of bias is particularly insidious because it's both prevalent in common datasets and difficult to detect through traditional evaluation methods. Traditional bias mitigation techniques, which often focus on model-level interventions or rely on language models for debiasing, have proven inadequate for addressing these spurious correlations, making new approaches like knowledge graph-based solutions particularly valuable. The authors argue that representing the dataset as a knowledge graph of object co-occurrences allows for a structured and controllable method to diagnose and mitigate these spurious correlations. Concept co-occurrence Biases in visual datasets (ConBias) To address this gap, the paper introduces ConBias, a novel framework designed to diagnose and mitigate concept co-occurrence biases in visual datasets. ConBias achieves this by representing visual datasets as knowledge graphs of concepts. This representation allows for a detailed examination of spurious concept co-occurrences and the identification of concept imbalances throughout the dataset. The problem setting for ConBias is as follows: Input: A biased visual dataset with images, corresponding class labels, and a concept set describing unique objects in the data. Challenge: Identify and mitigate object co-occurrence biases in the dataset, manifesting as spurious correlations between class labels and concepts. Goal: Generate an debiased augmented dataset with respect to the concepts and their corresponding class labels, enabling improved classifier training and generalization performance. Three Steps of ConBias As mentioned before, object co-occurrence bias refers to the spurious correlation between a label and an object causally unrelated to that label. By zeroing in on object co-occurrence bias, the authors aim to tackle a specific and prevalent type of bias that poses a significant challenge in visual recognition tasks. The ConBias framework specifically addresses this issue. The authors argue that generating images with a more uniform concept distribution across classes will improve the generalization capabilities of a classifier trained on the debiased dataset. At a high level, the ConBias method involves three steps: Concept Graph Construction maps the co-occurrence relationships between objects (concepts) and class labels. Concept Diagnosis uncovers imbalanced concept combinations across classes, pinpointing potential areas of object co-occurrence bias. Concept Debiasing rectifies these imbalances by generating new images with under-represented concept combinations. Let's go into these steps in more detail. Step 1: Concept Graph Construction The first step involves building a knowledge graph representing the visual dataset. This graph consists of nodes and edges, with weights assigned to the edges. The nodes in the graph represent the union of the dataset's class labels and the concept set describing unique objects present in the data. In addition to the class label, the concept set might consist of objects like "alley," "crosswalk," or "gas station" present in the images. The edges connect nodes that co-occur in the same image. For example, if an image contains both a "landbird" (class label) and a "tree" (concept), there would be an edge between the corresponding nodes. The weight of each edge signifies the frequency of co-occurrence between the two connected nodes. So, a higher weight indicates a more frequent co-occurrence of those particular concepts or class labels within the dataset. This step transforms the visual data into a structured graph format that captures the dataset's co-occurrence relationships between various concepts and class labels. Step 2: Concept Diagnosis Once the concept graph is constructed, ConBias analyzes it for concept imbalances, which may indicate biases in the original dataset. The framework achieves this through a series of definitions and operations: Class Clique Sets: For every class in the dataset, ConBias identifies groups of interconnected concepts (cliques) within the graph. Each clique represents a specific combination of concepts. These cliques are categorized based on their size, denoted by k, which refers to the number of concepts within each group. For instance, a 2-clique would represent a pair of co-occurring concepts, while a 3-clique would involve three. The framework constructs these class clique sets for every class in the dataset, considering various clique sizes ranging from 1 to the largest clique containing the specific class. Common Class Clique Sets: Next, ConBias focuses on the cliques shared across all classes in the dataset. These common cliques, denoted as K, are particularly important for bias analysis as they represent concept combinations appearing across different classes, enabling comparison of their co-occurrence frequencies. Imbalanced Common Cliques: This step identifies common cliques that exhibit uneven co-occurrence patterns across different classes, suggesting a potential bias. For each common clique, ConBias calculates the difference in co-occurrence frequencies between all class pairs. Larger discrepancies indicate a greater imbalance, suggesting a stronger potential bias. For example, in the Waterbirds dataset, the concept combination (Beach, Ocean) might be significantly more frequent in images labeled as "Waterbird" compared to "Landbird", revealing a potential bias. This analysis highlights concept combinations that exhibit imbalanced distributions despite being common across classes, suggesting the possibility of spurious correlations between concepts and class labels. Step 3: Concept Debiasing The final step of ConBias addresses the imbalances identified in the previous stage by generating new images that contain under-represented concept combinations. Based on the Concept Diagnosis step analysis, ConBias pinpoints the concept combinations and corresponding classes that require rebalancing. For each under-represented combination, the framework calculates the number of new images needed to achieve a balanced representation. ConBias creates these new images using a text-to-image generative model, such as Stable Diffusion. The prompts given to the model are descriptive phrases constructed from the underrepresented concept combinations. For example, if the combination "Waterbird, Tree" was underrepresented, ConBias would prompt the model with phrases like, "An image of a waterbird and a tree." To maintain the integrity of the original images and avoid unwanted modifications to the objects themselves, ConBias uses an inpainting technique. This technique involves generating the background with the desired concepts and then seamlessly inserting the original object into the scene, using ground-truth masks if available. These newly generated images containing balanced representations of the previously imbalanced concepts are then added to the original dataset to create an augmented dataset. This augmented dataset can then retrain the classifier, ideally reducing the impact of the identified biases on its performance. Key Findings Across multiple datasets (Waterbirds, UrbanCars, and COCO-GB), ConBias consistently demonstrated a substantial improvement in the generalization performance of classifiers. Specifically, it achieved notable gains in accuracy on both class-balanced and out-of-distribution (OOD) test sets. This improvement indicates that models trained on the debiased dataset generated by ConBias are better equipped to handle data that doesn't conform to the biases present in the original training data. Superiority over Traditional Augmentation Techniques: Compared to standard data augmentation methods like CutMix and RandAug, ConBias exhibited a marked advantage in mitigating object co-occurrence bias. This suggests that ConBias's targeted approach, which focuses on identifying and rectifying specific concept imbalances, is more effective than generic augmentation techniques that don't explicitly address bias. Outperformance of State-of-the-art Methods: ConBias outperformed ALIA. This recently proposed state-of-the-art method uses large language models for data debiasing. The authors attribute this superior performance to ConBias directly diagnosing and addressing biases within the dataset instead of relying on external language models that may introduce new biases. Effectiveness of Clique-Based Concept Balancing: The study found that leveraging the graph structure and employing clique-based concept balancing is crucial for effective bias mitigation. This approach, which analyzes concept co-occurrences within cliques of varying sizes, allows ConBias to identify and rectify more complex and subtle biases compared to simply examining single-concept frequencies. Importance of Concept Imbalance Discovery: ConBias's ability to identify concept imbalances significantly impacted its success. By pinpointing specific concept combinations over- or under-represented for certain classes, ConBias can guide the generation of synthetic data that effectively addresses these imbalances. The researchers note that this targeted approach is more effective than relying on diverse prompts from large language models, as done in ALIA. Benefits of Inpainting-Based Image Generation: Using an inpainting-based image generation method, which preserves the original object while modifying the background, proved beneficial for debiasing. This approach ensures that the synthetic data remains relevant to the classification task and avoids introducing artifacts that could hinder model training. These findings highlight the effectiveness of the ConBias framework in diagnosing and mitigating object co-occurrence bias in visual datasets, leading to improved model generalization and more reliable deep learning applications. They also underscore the importance of addressing bias directly within datasets and utilizing targeted approaches for data debiasing. By the way, come say "Hi!" if you're at NeurIPS 2024 in Vancouver! I'll be at the Voxel51 Booth (booth 415) - just follow the orange, you can't miss us!

https://medium.com/voxel51/using-knowledge-graphs-to-diagnose-and-debias-visual-datasets-31c464ded5fc

datascienceharp

A Data-Centric Look at Curation Strategies for Image Classification

Review of a Data-Centric AI Paper from NeurIPS 2024 —SELECT: A Large-Scale Benchmark of Data Curation Strategies for Image Classification

A Data-Centric Look at Curation Strategies for Image Classification Review of a Data-Centric AI Paper from NeurIPS 2024 -SELECT: A Large-Scale Benchmark of Data Curation Strategies for Image Classification This post is part of a five-part series examining notable data-centric AI papers from NeurIPS 2024. For brief summaries of all five papers, checkout my overview post, where you'll find links to each detailed analysis. While much of the machine learning discourse focuses on models, algorithms, and architectures, the critical role of data curation often remains in the shadows. Data curation, which involves carefully selecting and organizing data to create a dataset, profoundly impacts the performance and robustness of machine learning models, particularly in image classification. Despite growing awareness of data curation's significance, many studies fall short of best practices, often providing minimal information about their training data and its curation process. This lack of transparency obscures the vital connection between data quality and model performance, hindering progress toward more robust and reliable machine learning systems. Relevant Links: GitHub Repo Project page arXiv abstract Reviewer comments on OpenReview While data curation has historically been an implicit consideration in machine learning research, it's recently gained prominence as a research topic in its own right. In this paper, the authors bring data curation into sharper focus and establish it as a distinct research area by formalizing the task as a rational choice problem whose goal is maximizing the utility of the resulting dataset within specific cost constraints. The paper formalizes the task of data curation strategy as a function that takes a cost input and produces a set of samples drawn from a distribution over a set of plausible images. In data curation, costs can arise from various sources: Data Acquisition: Gathering images or image-text pairs from the web, specialized databases, or through synthetic generation can be computationally expensive and time-consuming. Labeling: Obtaining accurate labels for the data, whether through expert annotation, crowdsourcing, or automated methods, incurs costs. Filtering: Selecting the most informative and relevant samples from a large pool of data often requires human effort or sophisticated algorithms, both of which have associated costs. Data curation strategies can be viewed as a series of choices by curators to maximize the dataset's utility within a given cost constraint. In this paper, the authors discuss five data curation strategies: Expert Curation: Considered the gold standard, this strategy involves human-in-the-loop at all stages, including selecting the label set, prefiltering images, and assigning labels with expert oversight. This method results in high-quality datasets, such as the original ImageNet, but is costly due to extensive human effort. Crowdsourced Labeling: This approach reduces labelling costs using a wider pool of annotators. Experts define the label set, but image prefiltering is omitted. Annotators can apply multiple labels per image, potentially leading to class imbalances. Schema Matching: This strategy leverages the existence of well-curated datasets by mapping their label sets to create new datasets. A schema is created to connect labels across datasets, often requiring expert input. While schema creation is relatively low-cost, the quality and balance of the resulting dataset depend heavily on the source datasets. Synthetic Data Generation: This strategy bypasses the need for real images by using generative models to create synthetic images and labels. The models are trained on existing datasets and can be conditioned on various factors, such as label sets, text captions, or images. However, synthetic images often lack fidelity compared to real images, presenting a challenge for this approach. Embedding-Based Search: This method utilizes pre-trained computer vision models, often vision-language models like CLIP, to search large, unlabeled datasets for images relevant to target classes. This technique can efficiently retrieve images semantically similar to those in a reference dataset or matching specific text prompts by comparing image embeddings. However, this approach can introduce label noise, requiring further filtering or correction techniques. The underlying principle is that curators make rational choices to select a curation strategy that aims to maximize the utility of the set of samples while staying within the given cost constraint. Increasing the allowed cost typically allows for a larger and potentially more diverse set of samples, which is expected to lead to higher utility. In essence, their formalization casts data curation as an optimization problem: Objective: Maximize the utility of the curated dataset. Decision Variables: The curation strategy which encompasses choices about data sources, labeling methods, filtering techniques, and more. Constraint: The total cost of curation must not exceed the allowed budget. Curators must carefully consider the costs and benefits of different approaches to arrive at a dataset that effectively balances utility and resource constraints. Exploring Utility and Analytic Metrics in Data Curation The core idea is that a dataset possesses a certain level of utility, which reflects its effectiveness for the intended task (in this paper, the focus is image classification). This utility can be quantified through various metrics, broadly grouped into two categories: utility and analytic metrics. These metrics play distinct but complementary roles in assessing the effectiveness of different curation methods. Utility Metrics: Measuring Dataset Usefulness Through Model Training Utility metrics focus on measuring the practical usefulness of a curated dataset for training image classification models. They involve training models on the dataset and evaluating their performance on various tasks. The paper discussed the following key utility metrics: Base Accuracy: This metric measures the model's performance on a holdout set drawn from the same distribution as the baseline dataset. In this research, the baseline is the original ImageNet-train dataset, and the holdout set is ImageNet-val. Base accuracy directly measures how well a model trained on a particular dataset generalizes to unseen data from the same distribution. OOD Robustness: This metric assesses the model's ability to generalize to out-of-distribution (OOD) datasets, which differ in some way from the training distribution. This includes synthetic OOD shifts (e.g., ImageNet-C, which introduces image corruptions) and natural OOD shifts (e.g., ImageNet-Sketch, which uses sketches of objects). OOD robustness is crucial for evaluating a model's ability to handle real-world scenarios where the data may not perfectly match the training distribution. Fine-tuning: This metric evaluates the model's ability to adapt to new, unseen tasks after being pretrained on the curated dataset. Strong fine-tuning performance indicates that the pretrained model has learned generalizable features that transfer well to new domains. Self-Supervised Guidance: This metric uses a self-supervised learning method (specifically, DINO) to pretrain a model on the curated dataset without using any labels. The pretrained model is then evaluated on the ImageNet-val test set using k-NN classification. This approach measures the dataset's usefulness for learning representations without relying on explicit labels. Analytic Metrics: Characterizing Datasets Without Training In contrast to utility metrics, analytical metrics aim to capture the essential characteristics of a dataset without requiring model training. They offer insights into potential factors influencing model performance and can be used for rapid evaluation and comparison of different datasets. The paper categorizes analytic metrics as follows: Summary Statistics These metrics provide a basic overview of the dataset, including: Dataset Size: The number of unique samples in the dataset. Class Coverage: The number of classes in the label set represented in the dataset. Imbalance Metrics: These metrics capture the distribution of samples across classes, highlighting potential issues with class imbalance. The authors introduce two specific metrics: Left-Skewedness: This measures the concentration of samples in a few dominant classes. High left-skewedness indicates that a small number of classes account for a large proportion of the samples, which can bias the model towards those classes. Long-tailedness: This measures the proportion of classes with very few samples. A highly long-tailed dataset has many classes with limited representation, making it difficult for the model to learn effectively on those classes. Quality Metrics These metrics aim to assess the quality of the images and labels in the dataset. The sources consider several metrics, including: CLIPScore: This metric uses a CLIP model to evaluate the similarity between the images and their corresponding text labels, measuring image and label quality. CLIP-IQA: This metric uses CLIP and generic semantic opposite pairs (e.g., "good/bad", "bright/dark") to assess the quality of the images alone. Inception Score: This widely used metric measures the diversity and recognizability of generated images using a pretrained Inception v3 model. CMMD (CLIP Maximum Mean Discrepancy): This recent metric utilizes richer CLIP embeddings and the maximum mean discrepancy distance to evaluate image quality. Correlational Metrics These metrics examine the relationships between various dataset properties, such as: Correlation between precision and class count (indicating potential label noise in larger classes). Correlation between accuracy and confusion skewness (how concentrated model errors are on certain classes). Correlation between the accuracy of the ImageNet-1k model and the model trained on the shift dataset. Correlation between precision and recall. Correlation between class availability in ImageNet-1k and the shift dataset. In essence, utility metrics answer the "what" question (which datasets lead to better model performance), while analytic metrics help to answer the "why" question (what characteristics of the datasets contribute to those performance differences). The Focus is on Benchmarking, Not New Curation Methods While the paper introduces a framework for evaluating data curation strategies, it doesn't propose a novel method for data curation itself. The primary goal of the paper is to: Bring attention to the importance of data curation. Establish a standardized way to assess and compare different data curation strategies. Provide insights into the strengths and limitations of existing curation methods. The paper does, however, introduce SELECT and IMAGENET++. SELECT is a benchmark, criteria, and metrics to evaluate data curation strategies. IMAGENET++ is a dataset, a collection of curated image sets, used to test and compare different curation strategies using the SELECT benchmark. While closely related, they serve distinct purposes. SELECT: A Framework for Evaluation Purpose: Provide a standardized and comprehensive way to assess the quality and utility of datasets created using various data curation methods. It helps researchers compare different approaches and understand their strengths and weaknesses. Focus: Evaluate how well datasets support efficient learning for image classification tasks. Metrics: The utility and analytic metrics discussed above. Goal: Encourage a more systematic and rigorous evaluation of data curation strategies, moving beyond relying solely on base accuracy. Providing a diverse set of metrics highlights the importance of considering factors like robustness, generalization, and dataset properties when assessing the quality of curated data. IMAGENET++: A Testbed for Data Curation Strategies Purpose: A large-scale dataset specifically designed to evaluate the SELECT benchmark. It provides a collection of datasets (referred to as "shifts"), each curated using one of the five strategies previously discussed, allowing for direct comparison of their performance. Focus: Image classification, building upon the widely studied ImageNet dataset. Datasets: It includes the original ImageNet training set (the baseline representing expert curation) and five shifts (refer to Table 1 in the paper). Goal: Enable researchers to empirically test and compare different data curation strategies in a controlled setting. By training models on these shifts and evaluating them using the SELECT benchmark, the authors gain insights into how various curation methods impact model performance across various tasks and metrics. Key Findings The central finding of the research is that while no single reduced-cost data curation strategy outperforms the original expert-curated ImageNet dataset across all metrics, some methods, particularly embedding-based search techniques, exhibit promising results and are worthy of further exploration and refinement. Here are the key takeaways regarding the performance of each curation strategy: Expert Curation: As expected, the original ImageNet dataset, created through meticulous human effort, remains the best-performing dataset across most utility metrics, highlighting the enduring value of expert knowledge in crafting high-quality datasets. Crowdsourced Labeling: While less expensive than expert curation, this approach results in significant class imbalance, leading to subpar performance on most tasks. Surprisingly, even with human annotators, this method often underperforms compared to some less expensive methods. Embedding-Based Search: This strategy, utilizing CLIP embeddings to select images, emerges as the most promising reduced-cost method, consistently outperforming other techniques like synthetic image generation. However, it suffers from label noise, which hinders its ability to fully match expert-curated datasets Synthetic Data Generation: While offering potential cost savings, this method, relying on Stable Diffusion, struggles to generate high-quality images to compete with real image datasets. They note that current image quality metrics fail to accurately predict the utility of synthetic datasets, suggesting a need for better evaluation tools for this approach. Ultimately, choosing a data curation strategy involves a trade-off between cost and performance. This research underscores data curation's critical role in the success of machine learning models. The authors argue that continued research and development of more efficient and effective data curation strategies are crucial for unlocking the full potential of machine learning across various domains and applications. By the way, come say "Hi!" if you're at NeurIPS 2024 in Vancouver! I'll be at the Voxel51 Booth (booth 415) - just follow the orange, you can't miss us!

https://medium.com/voxel51/a-data-centric-look-at-curation-strategies-for-image-classification-b406f0436cff

datascienceharp

Are We Measuring What We Think We Are? The Perils of Contaminated Benchmark Datasets

Review of a Data-Centric AI Paper from NeurIPS 2024 — Intrinsic Self-Supervision for Data Quality Audits

Are We Measuring What We Think We Are? The Perils of Contaminated Benchmark Datasets Review of a Data-Centric AI Paper from NeurIPS 2024 - Intrinsic Self-Supervision for Data Quality Audits This post is part of a five-part series examining notable data-centric AI papers from NeurIPS 2024. For brief summaries of all five papers, checkout my overview post, where you'll find links to each detailed analysis. Machine learning has made incredible progress in recent years, with deep learning models achieving impressive results on various tasks. However, this progress is often measured and compared using benchmark datasets, which are supposed to be reliable and representative collections of data. But what happens when these very benchmarks are contaminated with errors? It turns out, yhis contamination seriously threatens the reliability of benchmark results, potentially leading to overestimating model performance and hindering scientific progress. Relevant links: GitHub Repo Project page arXiv abstract Reviewer comments on OpenReview Data cleaning in deep learning is especially important in low-data regimes, where poor-quality samples can substantially affect model performance. In this paper, the researchers aim to address the tension between the need to clean benchmark datasets and the principle of avoiding manipulating evaluation data. Several factors contribute to the authors' research question. First, they recognize the limitations of traditional data-cleaning approaches in handling the massive scale of modern datasets, particularly in computer vision. Second, the widespread use of benchmark datasets for performance comparisons necessitates a robust and efficient method for identifying and mitigating data quality issues to ensure reliable evaluations. Finally, the authors acknowledge the prevalence of data quality issues, even in curated medical datasets, highlighting the importance of data-centric AI principles for improving the reliability of machine learning models. They point out that the focus on data quantity over quality has led to varying noise levels in datasets. They also acknowledge the heavy reliance on benchmark datasets by deep learning practitioners despite them being known to contain data quality issues that can lead to over-optimistic results. The need for comparable results has driven the reliance on benchmarks despite their limitations, creating a tension between the need for data cleaning and the avoidance of manipulating evaluation data. Data Quality Issues in Benchmark Datasets The paper highlights three specific types of data quality issues: Off-topic samples: These are images mistakenly included in the dataset, deviating from the intended data distribution. Examples include images from unrelated domains or those affected by device malfunctions. The presence of such samples adds noise to evaluation metrics and can confuse the training process. Near duplicates: These images depict the same object, potentially resulting in data leaks between training and evaluation sets, reducing training variability, and leading to over-optimistic performance estimations. Label errors: Incorrectly annotated samples can poison the training process and result in inaccurate evaluation. As mentioned in the paper, there are several reasons why these data quality issues arise: Limited Manual Curation: While some datasets undergo rigorous manual verification, others rely on less intensive curation processes, potentially leading to a higher prevalence of errors. Data Acquisition Challenges: Collecting and assembling large datasets can introduce errors, such as accidentally including thumbnail images or failing to track metadata that links images with a common origin properly. Annotation Complexities: Annotating images, especially for fine-grained tasks or in specialized domains like medicine, can be challenging, leading to occasional mislabeling To address these data quality issues, the authors propose SELFCLEAN. SELFCLEAN: A Data Cleaning Method SELFCLEAN addresses the problem of data quality issues within benchmark datasets by employing a two-step process: Representation Learning: The first step involves training a deep feature extractor using self-supervised learning (SSL) on the dataset to be cleaned. This training is performed on the entire dataset, encompassing any existing data quality issues. The authors suggest using either SimCLR or DINO as the SSL objective. Both have demonstrated success in learning meaningful representations even from noisy data. The learned representations are then compared using cosine similarity and the associated distance, scaled to a range of [0,1]. Using SSL for representation learning ensures that SELFCLEAN does not inherit any annotation biases, as it relies solely on the data's inherent structure. Distance-Based Indicators: Once the feature extractor is trained, SELFCLEAN utilizes distance-based indicators to identify potential data quality issues. These indicators are used with the learned representations to identify candidate data quality issues based on distances between samples in the latent space. Each data quality issue type is addressed with a specific indicator function that leverages the local structure of the embedding space: Off-topic samples: These are identified using agglomerative clustering with single linkage. Samples that merge later with larger clusters during the clustering process, indicating greater distance from the main data clusters in the latent space, are flagged as potential off-topic samples. Near duplicates: SELFCLEAN computes pairwise distances between all samples in the latent space to detect near duplicates. Pairs with very small distances, suggesting high similarity in their representations, are flagged as potential duplicates. Label errors: SELFCLEAN assesses the consistency of labels by comparing distances between a sample and its nearest neighbors from the same and different classes. If a sample exhibits a significantly smaller distance to neighbours from a different class than neighbours from its class, it raises suspicion of a potential label error. SELFCLEAN Operating Modes SELFCLEAN offers two modes of operation, allowing users to choose between a fully automated approach and a method incorporating human oversight: human-in-the-loop and fully automatic. Human-in-the-Loop In this mode, SELFCLEAN produces a ranked list of potential data quality issues. A human curator then inspects the top-ranked samples, confirming and correcting the identified problems or determining a suitable rank threshold to achieve a desired balance between precision and recall. This mode acknowledges the limitations of fully automated cleaning, especially for complex cases requiring subjective judgment, and the possibility that some flagged samples may be valuable edge cases rather than errors. Fully Automatic This mode uses the score distributions generated by the distance-based indicators. Typically, these scores exhibit a smooth distribution for clean samples, with contaminated samples receiving significantly lower scores, separating them from the bulk of the data. This allows for the automated identification of problematic samples based on statistically determined outlier thresholds. The method utilizes two hyperparameters: a "contamination rate guess" representing an estimated upper bound on the fraction of issues in the dataset, and a "significance level" defining the desired statistical confidence in the outlier detection process. Experiments to Evaluate SELFCLEAN The authors conduct a series of experiments to evaluate the effectiveness of SELFCLEAN in detecting off-topic samples, near duplicates, and label errors. These experiments involve synthetic and natural contamination in various benchmark datasets. Datasets: The authors utilize twelve datasets encompassing general vision and medical imaging. The general vision benchmarks include ImageNet, STL-10, CelebA, and Food-101N. The medical datasets consist of CheXpert, VinDr-BodyPartXR, PatchCamelyon, HAM10000, ISIC-2019, Fitzpatrick17k, DDI, and PAD-UFES-20, covering modalities such as X-ray, histopathology, and dermatoscopy images. Evaluation metrics: Performance evaluation relies on ranking metrics, specifically AUROC (Area Under the Receiver Operating Characteristic curve) and AP (Average Precision). AUROC measures the likelihood of a relevant sample being ranked higher than an irrelevant one. At the same time, AP assesses precision across all recall values, accounting for the proportion of positive and negative samples. Synthetic Contamination Experiments To compare SELFCLEAN with competing methods, the authors create synthetic datasets by introducing specific types of contamination into benchmark datasets (STL-10, VinDr-BodyPartXR, and DDI). In this experiment, they used the following contamination strategies: Off-topic samples: Adding images from a different category or applying Gaussian blurring. Near duplicates: Augmenting existing images with transformations like rotation, flipping, and blurring or creating collages with artifacts. Label errors: Randomly changing labels uniformly or proportionally to class prevalence. Natural Contamination Experiments The authors evaluate cleaning on naturally occurring data quality issues in benchmark datasets (ImageNet, Food-101N, CelebA, HAM10000, ISIC-2019, PAD-UFES-20, and CheXpert). They design two experiments: Metadata comparison: Measuring how well SELFCLEAN's ranking aligns with available metadata. For example, they use labels in CelebA that indicate images featuring the same celebrity and metadata in HAM10000 and ISIC-2019 that link images depicting the same skin lesion. Human annotator comparison: Comparing SELFCLEAN rankings against human verification for the top-ranked images. Key Findings SELFCLEAN with DINO pre-training consistently outperforms competing approaches across all three types of data quality issues (off-topic samples, near duplicates, and label errors) in terms of both AUROC and AP. Dataset-specific representations learned through SSL tend to outperform general-purpose representations, emphasizing the significance of capturing the dataset's context for effective data cleaning. Human annotators confirm significantly more data quality issues in the top-ranked images identified by SELFCLEAN compared to random samples, demonstrating that the method's rankings align well with human assessment. The effectiveness of SELFCLEAN heavily relies on pre-training the deep feature extractor using SSL. The choice of SSL objective and dataset used for pre-training significantly influences the results. DINO generally yields the best performance. Removing problematic samples identified by SELFCLEAN leads to significant changes in downstream classification performance, highlighting the practical importance of data cleaning for accurate model evaluation. Applying SELFCLEAN to benchmark datasets reveals the presence of various data quality issues, emphasizing the need for data cleaning and raising concerns about the reliability of reported results. The findings from the SELFCLEAN methodology expose the widespread presence of data quality issues like off-topic samples, near duplicates, and label errors, even in well-regarded benchmark datasets. Importantly, the paper underscores that identifying these issues isn't simply about discarding "bad" data but gaining a deeper understanding of the dataset's composition and potential biases, ultimately leading to more informed model development and interpretation. Instead, they advocate for a more comprehensive approach to data quality: Understanding Dataset Composition: Identifying relationships between samples, such as near duplicates even within the same data split, can provide valuable insights into the dataset's structure and potential biases. Enhancing Model Robustness: Cleaning the training data can lead to more robust models that are less susceptible to the negative effects of noisy data. Building Trust in Benchmarks: By addressing data quality issues, we can restore confidence in benchmark results and ensure that they accurately reflect the progress of machine learning By the way, come say "Hi!" if you're at NeurIPS 2024 in Vancouver! I'll be at the Voxel51 Booth (booth 415) - just follow the orange, you can't miss us!

https://medium.com/voxel51/are-we-measuring-what-we-think-we-are-the-perils-of-contaminated-benchmark-datasets-037c61932d82

datascienceharp

CoTracker3: Enhanced Point Tracking with Less Data

A new semi-supervised approach achieves state-of-the-art performance with a thousandfold reduction in real data requirements.

CoTracker3: Enhanced Point Tracking with Less Data A new semi-supervised approach achieves state-of-the-art performance with a thousandfold reduction in real data requirements. Introduction to Point Tracking In computer vision, point tracking estimates the movement of specific points within a video over time. In this task, a model is trained to identify and maintain correspondences between tracked points across multiple frames in a video. However, accurately tracking points in videos isn't a trivial task. First, you have to worry about occlusions. This is a massive pain. I'd put this as the primary challenge for point trackers because objects in a scene can obstruct the view of tracked points, leading to inaccurate tracking. Second, a tracked point can leave the field of view. As objects and the camera move, tracked, points will go with them, often out of the camera's field of view. This makes it difficult for some trackers to predict their trajectory. Third, you need to consider the appearance of points. Lighting and perspective changes can affect how a point looks. Rapid movements can make tracking difficult. Points may appear larger or smaller as objects move closer or farther from the camera. Last but not least is computational complexity. Computational complexity becomes a bottleneck when tracking many points simultaneously, especially considering dependencies between tracks. The Applications of Point Tracking That Make it a Problem Worth Tackling Despite what some folks may think, computer vision is not a solved problem, and the point tracking task has various applications that push the progress in other areas of computer vision. For example: Motion Analysis: Understanding motion provides insights into object movement and scene dynamics, which is essential for action recognition and tracking. 3D Reconstruction: You can infer 3D information by tracking points across multiple scene views. Video Editing and Special Effects: You can use point tracking to stabilize shaky footage, insert objects seamlessly, and apply effects that track specific points of interest Robotics and Autonomous Navigation: This scenario requires real-time tracking to perceive and interact with the environment, enabling tasks like obstacle avoidance and object manipulation. A Brief History of Point Tracking Historically, there have been a few dominant paradigms for this task. Optical Flow Optical flow models estimate dense instantaneous motion. They can be classical approaches, deep learning methods, or transformer-based methods. Classical Approaches: Traditional optical flow methods relied on brightness constancy equations and often combined local and global flow estimations3. Deep Learning-Based Methods: FlowNet and DCFlow used convolutional neural networks (CNNs) for this task. RAFT and its Variants: RAFT was novel as it used incremental flow updates and 4D cost volumes. This model inspired several follow-ups that further improved accuracy and efficiency. Transformer-Based Methods: Of course, in recent years, Transformers have been all the rage. For example, FlowFormer tokenizes the 4D cost volume, and GMFlow utilizes a softmax with self-attention for refinement. There's also Perceiver IO, which proposed a unified transformer architecture for various tasks, including optical flow. Optical flow methods are powerful for estimating dense instantaneous motion but are not ideal for long-term point tracking due to error accumulation. Multi-Frame Optical Flow Multi-frame optical flow models extend optical flow to multiple frames but are still not designed for long-term tracking or occlusion handling. While initial attempts to extend optical flow to multiple frames relied on Kalman filtering for temporal consistency, more recent approaches include: Modern Dense Flow Methods: Recent multi-frame optical flow models generate dense fields. RAFT can be adapted for multi-frame estimation through a warm-start approach. VideoFlow explicitly integrates forward and backward motion features across three to five consecutive frames to refine flow estimates. Multi-Flow Dense Tracker (MFT): MFT estimates flow between distant frames and selects the most reliable chain of optical flows to ensure consistent tracking. The methods can estimate flow across multiple frames but are not designed for long-term point tracking and struggle with handling occlusions, especially when points are occluded for a long time. Point Tracking Point tracking models track sparse sets of points over time. Unlike optical flow, which estimates dense instantaneous motion, point tracking focuses on a sparse set of points and tries to maintain their correspondence across multiple frames. Some models that follow this paradigm include: Particle Video: Particle Video pioneered the Tracking Any Point (TAP) concept but was limited in handling occlusions. PIPs: PIPs, building upon Particle Video, introduced improvements to track points through occlusions more reliably. This model utilizes a sliding window approach and restarts tracks from the last visible frame of a point. TAP-Vid: TAP-Vid introduced a new benchmark and a simple baseline for TAP, and if there's anything new benchmarks do, it's pushing the field forward! TAPIR: Combining concepts from TAP-Vid and PIPs, TAPIR is a two-stage, feed-forward point tracker that significantly improves tracking performance, especially in occlusion handling. PIPs++: Addressing long-term tracking, PIPs++, a simplified version of PIPs, was introduced alongside a benchmark for long-term tracking. OmniMotion: OmniMotion optimizes a volumetric video representation and refines correspondences in a canonical space. It often achieves high accuracy but requires computationally expensive test-time optimization. Many point tracking models, like PIPs and PIPs++, track points independently; however, points in a video often exhibit strong dependencies, such as belonging to the same object. The original CoTracker model uses these dependencies by performing joint tracking of many points. CoTracker is a transformer-based model that tracks many 2D points in extended video sequences. Unlike previous methods that typically track points independently, CoTracker introduces the concept of joint tracking. This means paying attention to dependencies between tracked points, leading to enhanced accuracy and robustness, mainly when dealing with occlusions or points moving out of the camera view. Here are the new techniques that CoTracker developed to improve on previous methods: CoTracker uses a transformer architecture with an attention mechanism to share information between tracked points to better understand scene motion and predict occluded points. This process, known as joint point tracking, differs from methods that track points independently. CoTracker is one of the few that uses deep networks for joint tracking. Support Points improve tracking accuracy, especially for tracking a single point or a few points. Support points (though not explicitly requested by the user or provided as an argument) offer additional context to the model. Different configurations include "global" (a regular grid across the image) and "local" (a grid centered around the target point). Experiments show that combining global and local support points gives the best performance. CoTracker introduces proxy tokens to address the computational cost associated with attention when dealing with many tracks. The architecture represents tracks using a grid of tokens, with each token encoding a specific track's position, visibility, appearance, and correlation features at a given time. These tokens efficiently represent a subset of tracks, reducing memory complexity and enabling the model to jointly track a near-dense set of points on a single GPU during inference. This approach is like using registers to reduce memory complexity in other transformer architectures. While operating online with a sliding window approach, CoTracker uses unrolled training (borrowing the concept from recurrent networks). This method optimizes a network by unrolling its application over multiple overlapping windows during training. This improves long-term tracking capabilities, especially for occluded points. The windowed approach can process videos of arbitrary length by initializing subsequent windows with information from preceding ones, mimicking a recurrent network. However, despite its impressive performance, CoTracker has some limitations: CoTracker is trained on synthetic data, which makes generalizing to complex real-world scenes - with elements like reflections and shadows - challenging. The model is sensitive to discontinuous videos, and performance might degrade in videos with multiple shots or discontinuous segments, as it is primarily designed for continuous video sequences. Introducing CoTracker3 Building upon the foundation laid by CoTracker, CoTracker3 has a simpler architecture, improved data efficiency, and greater flexibility. While maintaining the core concept of joint point tracking, CoTracker3 refines and streamlines various aspects, achieving state-of-the-art results with significantly less training data. Here are some of the improvements and innovations introduced by CoTracker3 that I feel are interesting: Architectural Simplifications and Enhancements CoTracker3 uses a 4D correlation feature representation that captures spatial relationships between features in different frames (introduced in the LocoTrack paper). However, it simplifies the processing of these features by employing a straightforward multi-layer perceptron (MLP) instead of LocoTrack's more complex ad-hoc module. This reduces computational overhead and maintains representational power. In CoTracker, visibility flags - indicating whether a point is visible or occluded - were updated by a separate network. CoTracker3 integrates this process directly into the main transformer, updating visibility flags alongside other track attributes at each iteration. This simplifies the architecture and improves efficiency. CoTracker3 has online and offline versions with the same architecture but different training procedures. The online version operates in a sliding window for real-time tracking, while the offline version processes the entire video for improved bi-directional tracking and handling of occlusions. Semi-Supervised Training Pipeline CoTracker3 uses a semi-supervised training strategy using real-world videos without manual annotations. It uses multiple existing point trackers trained on synthetic data to generate pseudo-labels for real videos and then trains a student model on this larger, pseudo-labeled dataset. This approach is data efficient. It outperforms BootsTAPIR - a state-of-the-art tracker trained on 15 million real videos - using only 15,000 real videos, a whopping thousandfold decrease in data requirements. CoTracker3 uses SIFT feature detection to guide the selection of query points for pseudo-labeling, prioritizing points deemed "good to track." This improves pseudo-labeled data quality and training stability. Performance Improvements CoTracker3 consistently outperforms existing trackers on various benchmarks, including TAP-Vid, Dynamic Replica, and RoboTAP. It does an excellent job tracking occluded points, especially in offline mode. This is thanks to its joint tracking capability and access to the entire video sequence. CoTracker3's architectural simplifications result in a leaner model that runs 27% faster than LocoTrack, the previous fastest point tracker, despite incorporating cross-track attention for joint tracking. Experiments with increasing amounts of pseudo-labeled data demonstrate that CoTracker3 continues to improve with more real-world data. CoTracker3 also benefits from self-training, where it is fine-tuned on its own predictions on real videos. This bridges the gap between synthetic training data and real-world scenarios. Limitations Despite its advancements, CoTracker3, like its predecessor, has its limitations: The pseudo-labeling pipeline's performance depends on the quality and diversity of the chosen teacher models. As the student model approaches the performance of its teachers, its ability to improve will plateau, necessitating the introduction of stronger or more diverse teachers for continued advancement. Ensure diversity and representativeness in the pseudo-labeled data to address the potential for overfitting to specific domains with pseudo-labeling. CoTracker and CoTracker3 are significant point-tracking advancements, mainly when dealing with long video sequences and challenges like occlusions. CoTracker introduced the innovative concept of joint tracking through a transformer network, achieving state-of-the-art performance but relying heavily on synthetic training data. CoTracker3 builds upon this foundation, introducing architectural simplifications, a novel semi-supervised training pipeline, and improved efficiency. By leveraging multiple pre-trained trackers as teachers to generate pseudo-labels for real-world videos, CoTracker3 significantly reduces the dependency on synthetic data. It achieves even better accuracy with a thousandfold reduction in real data requirements. Both models highlight the power of considering dependencies between tracked points and utilizing context to enhance tracking accuracy. Next Steps Now that you're up to speed on the task of point tracking and the CoTracker3 model, it's time to get hands on with some code! Check out this blog where you'll learn about: Online vs. Offline Modes: Distinguish between CoTracker3's online (real-time, forward-only) and offline (bidirectional, better accuracy but memory-intensive) modes. Running Inference: Learn how to download a pre-trained CoTracker3 model and run inference on video data using PyTorch. Visualizing Results: See how to visualize the tracked points and their visibility over time. FiftyOne Integration: Understand how to parse and integrate CoTracker3's output into FiftyOne, a powerful dataset visualization and analysis tool, for exploring and interacting with the tracking results. Memory Management: Learn practical tips for managing GPU memory when working with large video files, including pre-processing techniques like frame sampling and rate reduction.

https://medium.com/voxel51/cotracker3-enhanced-point-tracking-with-less-data-00f51eb23110

datascienceharp

3D Scene Understanding: Open3DSG’s Open-Vocabulary Approach to Point Clouds

A CVPR Paper Review and Cliff’s Notes

3D Scene Understanding: Open3DSG's Open-Vocabulary Approach to Point Clouds A CVPR Paper Review and Cliff's Notes Understanding 3D environments is a critical challenge in computer vision, particularly for robotics and indoor applications. The paper, Open3DSG: Open-Vocabulary 3D Scene Graphs from Point Clouds with Queryable Objects and Open-Set Relationships, introduces a novel approach for predicting 3D scene graphs from point clouds in an open-world setting. The paper's main contribution is a method that leverages features from powerful 2D vision language models (VLMs) and large language models (LLMs) to predict 3D scene graphs in a zero-shot manner. This allows for querying object classes from an open vocabulary and predicting inter-object relationships beyond a predefined label set. This research moves beyond traditional, predefined class limitations by leveraging vision-language models to identify and describe arbitrary objects and their relationships, setting a new standard for machine perception and interaction in complex environments. The Problem Current 3D scene graph prediction methods depend heavily on labeled datasets, restricting them to a fixed set of object classes and relationship categories. This limitation reduces their effectiveness in real-world applications where a broader and more flexible vocabulary is necessary. Insufficiencies of Current Methods Fixed Label Set: Traditional methods are confined to a narrow scope of training data, hindering their ability to generalize to unseen object classes and relationships. Lack of Compositional Understanding: Existing 2D VLMs struggle with modeling complex relationships between objects, which is crucial for accurate 3D scene graph predictions. Inflexibility: Supervised training with fixed labels cannot adapt to new or rare object classes and relationships, limiting the practical utility of the models. The Solution The paper proposes Open3DSG, an approach to learning 3D scene graph prediction without relying on labelled scene graph data. The method co-embeds the features from a 3D scene graph prediction backbone with the feature space of open-world 2D VLMs. How the Solution Works Initial Graph Construction: The method begins by constructing an initial graph representation from a 3D point cloud using class-agnostic instance segmentation. Feature Extraction and Alignment: Features are extracted from the 3D scene using a Graph Neural Network (GNN) and aligned with 2D vision-language features. Object Class Prediction: At inference time, object classes are predicted by computing the cosine similarity between the distilled 3D features and open-vocabulary queries encoded by CLIP. Relationship Prediction: Inter-object relationships are predicted using a feature vector and the inferred object classes, providing context to a large language model. Improvements Introduced Open-Vocabulary Predictions: The method can predict arbitrary object classes and relationships, not limited to a predefined set. Zero-Shot Learning: This approach allows for zero-shot predictions. It can generalize to new objects and relationships without additional training data. Compositional Understanding: The method enhances the ability to model complex relationships between objects by combining VLMs with LLMs. Why It's Better Detail and Realism: The method provides fine-grained semantic descriptions of objects and relationships, capturing the complexity of real-world scenes. Efficiency: By aligning 3D features with 2D VLMs, the method achieves effective scene graph predictions without requiring extensive labeled datasets. Computational Power: The approach leverages powerful existing models (like CLIP and large language models), enhancing its ability to generalize and perform complex reasoning tasks. Key Contributions First Open-Vocabulary 3D Scene Graph Prediction: This paper presents the first method for predicting 3D scene graphs with an open vocabulary for objects and relationships. Integration of VLMs and LLMs: This approach combines the strengths of vision-language models and large language models to improve compositional understanding. Interactive Graph Representation: The method allows for querying objects and relationships in a scene during inference time. Results Experimental Validation: The method was tested on the closed-set benchmark 3DSSG, showing promising results in modelling compositional concepts. Comparison with State-of-the-Art Methods: Open3DSG demonstrated the ability to handle arbitrary object classes and complex inter-object relationships more effectively than existing methods. Final Thoughts As a forward-thinking system, Open3DSG's benefits are twofold: Enhances the expressiveness and adaptability of 3D scene graphs Paves the way for a more intuitive machine understanding of complex environments. With applications ranging from robotics to indoor scene analyses, the potential is vast. The improvements introduced by Open3DSG are significant as they enable a more flexible and detailed understanding of 3D scenes. This can be particularly important for computer vision and robotics applications, where understanding complex scenes is crucial. Will you be at CVPR 2024? Come by the Voxel51 booth and say "Hi!"!

https://medium.com/voxel51/3d-scene-understanding-open3dsgs-open-vocabulary-approach-to-point-clouds-69d443d29cb2

datascienceharp

SelfEQ Enhances Visual Grounding with Self-Consistency

A CVPR Paper Review and Cliff’s Notes

SelfEQ Enhances Visual Grounding with Self-Consistency A CVPR Paper Review and Cliff's Notes Precise visual grounding remains a challenging yet essential task, particularly when models encounter varied textual descriptions. The paper "Improved Visual Grounding through Self-Consistent Explanations" tackles this head-on by introducing a method that leverages paraphrases to enhance model consistency and localization accuracy without relying on extensive annotations. This approach promises significant improvements in aligning visual and textual data, making it a critical advancement for engineers focused on refining AI's interpretative capabilities. The main contribution is introducing a weakly-supervised strategy called Self-Consistency Equivalence Tuning (SelfEQ), which leverages paraphrases to consistently improve the model's ability to localize objects in images. The Problem Existing Challenge Vision-and-language models trained to match images with text often struggle with the precise localization of objects, especially when the textual descriptions vary slightly (e.g., "frisbee" vs. "disc"). The challenge is to improve these models' grounding abilities without relying on extensive object location annotations. Current Methods and Their Insufficiencies Current methods often require additional finetuning with a bounding box or segment annotations or depend on pretrained object detectors. These approaches are limited by their need for detailed annotations and can lack consistency when handling varied textual descriptions. Specific Issues Lack of Detail: Existing models may not handle diverse vocabulary well, leading to inconsistent localization. Inconsistency: Models may fail to provide consistent visual explanations for paraphrased textual inputs referring to the same object. The Solution The paper proposes SelfEQ, which encourages self-consistent visual explanations for paraphrased text inputs. This method involves generating paraphrases using a large language model and finetuning the vision-and-language model to ensure that the original and paraphrased texts map to the same image regions. How It Works - Start with an Existing Method: The ALBEF model, which aligns images and text using image-text pairs without object location annotations, serves as the foundation. Improvements by SelfEQ 1. Paraphrase Generation: A large language model (e.g., Vicuna-13B) is used to create paraphrases for text descriptions. 2. Self-Consistency Tuning: Finetunes the model using GradCAM to ensure consistent visual attention maps for original and paraphrased texts. Why It's Better Benefits of the New Approach Expanded Vocabulary: The model can handle a broader range of textual descriptions. Improved Localization: SelfEQ enhances the precision and consistency of object localization without requiring bounding box annotations. Efficiency: The approach leverages weak supervision, reducing the need for detailed annotations and making the finetuning process more efficient. Key Contributions - Novel Objective (SelfEQ): Introduces a self-consistency equivalence tuning objective to improve visual grounding. - Paraphrase Utilization: Employs large language models to generate high-quality paraphrases, enhancing the model's vocabulary handling. - Performance Improvements: Achieves significant improvements in standard benchmarks (Flickr30k, ReferIt, RefCOCO+). Results Testing and Performance The new method was tested on several benchmarks (Flickr30k, ReferIt, RefCOCO+), showing substantial improvements: Flickr30k: 84.07% (4.69% absolute improvement) ReferIt: 67.40% (7.68% absolute improvement) RefCOCO+: 75.10% (test A), 55.49% (test B) (3.74% average improvement) Comparison with State-of-the-Art SelfEQ outperforms several prior methods, especially those that do not use box annotations, demonstrating better localization performance and vocabulary handling. Final Thoughts The improvements presented in this paper enhance the robustness and applicability of vision-and-language models in visual grounding tasks. By focusing on self-consistent explanations and leveraging weak supervision, the authors provide a pathway for models to handle a wider range of textual inputs more effectively. This work is essential for advancing research in visual grounding and making models more adaptable to real-world scenarios. Learn more here: Paper Project page GitHub If you'll be at CVPR this year, be sure to come and say "Hi!"

https://medium.com/voxel51/selfeq-enhances-visual-grounding-with-self-consistency-cdaba01e236c

datascienceharp

CLAP: Enhancing Linear Probing for Efficient Few-Shot Learning in Vision-Language Models

A CVPR Paper Review and Cliff’s Notes

CLAP: Enhancing Linear Probing for Efficient Few-Shot Learning in Vision-Language Models A CVPR Paper Review and Cliff's Notes Few-shot learning has become increasingly important for adapting large pre-trained vision-language models (VLMs) like CLIP to downstream tasks with limited labelled data. However, current state-of-the-art methods for this efficient transfer learning (ETL) scenario often make unrealistic assumptions and require impractical per-task hyperparameter tuning. In their recent paper, "A Closer Look at the Few-Shot Adaptation of Large Vision-Language Models," the authors discuss these issues and propose a novel approach called CLAP (CLass-Adaptive linear Probe). CLAP outperforms existing methods across various benchmarks and operates under more realistic and practical constraints. In this post, I'll review the key insights from the paper, understand the limitations of current methods, and see how CLAP addresses these challenges to push the state-of-the-art in the few-shot adaptation of VLMs. The Problem Adapting large pre-trained VLMs to downstream tasks with only a few labelled examples is challenging. Current state-of-the-art ETL methods for this few-shot adaptation scenario have limitations: Existing methods make unrealistic assumptions about access to a large corpus of labeled data for hyperparameter tuning. They require carefully tuning hyperparameters for each specific task using a large validation set, which is unrealistic. The hyperparameters optimized for one task don't generalize well to other tasks. They can dramatically underperform simple zero-shot predictions in the presence of distribution shifts. The Solution The authors propose a new approach called CLAP: It builds on a designed Linear Probing (LP) baseline initialized with CLIP's zero-shot (ZS) class prototypes. This ZS-LP baseline already outperforms more complex ETL methods. To further improve ZS-LP, CLAP introduces a constrained optimization objective that penalizes large deviations of the learned class prototypes from the original zero-shot prototypes during adaptation. It uses an Augmented Lagrangian Multiplier method to optimize the constrained objective. The ALM method is adapted to use class-wise penalty multipliers (vs sample-wise) to handle class imbalances and work with data augmentation. CLAP has several advantages over existing ETL methods: It performs consistently across various tasks and datasets with the same configuration. No per-task hyperparameter tuning is needed. It substantially outperforms state-of-the-art ETL approaches in all evaluated scenarios. The efficient linear probing architecture enables faster adaptation with less computation. The realistic adaptation setting without relying on a large validation set makes it more practical. Key Contributions Empirically showing that SoTA ETL methods require unrealistic per-task hyperparameter tuning and can underperform simple baselines without it. Proposing CLAP, a principled approach to improve Linear Probing for few-shot adaptation of VLMs. It constrains deviations from zero-shot prototypes and eliminates hyperparameter tuning. Results Results show CLAP delivers consistent strong performance across tasks with a fixed configuration, substantially outperforming state-of-the-art ETL methods in all cases. CLAP is evaluated on: Few-shot adaptation on 11 classification datasets Domain generalization scenarios Comparison to full fine-tuning methods Ablation studies validating design choices Final Thoughts This work highlights key issues with current few-shot adaptation methods for VLMs. It proposes a novel principled approach to address them. CLAP's strong performance with a fixed configuration across various tasks and datasets demonstrates the importance of designing methods that can realistically adapt without relying on impractical hyperparameter tuning. The authors hope this moves the field towards more practical and robust ETL solutions. Learn more here: Paper on arXiv Project Page GitHub If you'll be at CVPR this year, be sure to come and say "Hi!"

https://datascienceharp.medium.com/clap-enhancing-linear-probing-for-efficient-few-shot-learning-in-vision-language-models-2681c01a699d

datascienceharp

Patch-wise Attention Enhances Fine-Grained Visual Recognition

A CVPR Paper Review and Cliff’s Notes

Patch-wise Attention Enhances Fine-Grained Visual Recognition A CVPR Paper Review and Cliff's Notes You don't usually think of two things in the same sentence: creepy crawlies and cutting-edge AI. However, this combination will improve agriculture because if we can accurately identify insect species, we can protect our crops and ensure food security. The paper "Insect-Foundation: A Foundation Model and Large-scale 1M Dataset for Visual Insect Understanding" buzzes into the world of precision agriculture, tackling the need for accurate insect detection and classification. It hatches a novel dataset, "Insect-1M," swarming with 1 million images of insects, each meticulously labelled with detailed taxonomic info. The Problem In precision agriculture, accurately identifying and classifying insects is crucial for maintaining crop health and ensuring high-quality yields. Existing methods face several challenges: Current insect datasets are significantly smaller and less diverse than needed. For instance, many datasets contain only tens of thousands of images and cover a limited number of species. Given the estimated 5.5 million insect species, this is inadequate, leading to poor generalization and coverage for practical applications. Existing datasets often fail to provide the fine-grained details to distinguish similar insect species. Many datasets lack multiple images per species, diverse angles, or high-resolution images that capture subtle, distinguishing features. This makes it difficult for models to differentiate between species with minor but crucial variations. Many datasets do not include comprehensive taxonomic hierarchy or detailed descriptions. They often provide basic labels without deeper taxonomic context, such as genus or family levels. This limits the models' ability to learn effectively, as they miss out on the rich relational information within the insect taxonomy. The Solution The authors propose two main contributions: the "Insect-1M" dataset and a new Insect Foundation Model. Insect-1M Dataset - Contains 1 million images spanning 34,212 species, significantly larger than previous datasets. - Includes six hierarchical taxonomic levels (Subphylum, Class, Order, Family, Genus, Species) and auxiliary levels like Subclass, Suborder, and Subfamily. - Provides detailed descriptions for each insect, enhancing the model's understanding and training. Insect Foundation Model The Insect Foundation Model is designed to overcome fine-grained insect classification and detection challenges. Here's a detailed overview of its components: Image Patching - Patch Extraction: Input images are divided into smaller patches, allowing the model to focus on localized regions of the image. - Patch Pool Creation: These patches form a pool the model uses for further processing. Patch-wise Relevant Attention - Relevance Scoring: Each patch is assigned a relevance score based on its importance for classification. This is done by comparing patches to masked images, highlighting subtle differences. - Attention Weights: Patches with higher relevance scores are given more attention, guiding the model to focus on the most informative parts of the image. Attention Pooling Module - Aggregation of Information: The attention pooling module aggregates information from the patches, using the attention weights to prioritize the most relevant features. - Feature Extraction: This process helps extract detailed and accurate features to distinguish similar insect species. Description Consistency Loss The model incorporates a description consistency loss, which aligns the visual features extracted from the patches with the textual descriptions of the insects. Text decoders contribute to the description consistency loss, which ensures that the visual and textual features are consistent and complementary. By minimizing this loss, the model enhances its understanding and classification accuracy. Text Decoders 1. Feature Extraction: The text decoders extract semantic features from the textual descriptions. These features encapsulate the essential information conveyed in the descriptions. 2. Alignment with Visual Features: The extracted textual features are aligned with the visual features obtained from the image patches. This alignment is facilitated through attention mechanisms, ensuring that the model learns to associate specific visual patterns with corresponding textual descriptions. Multimodal Text Decoders Multimodal text decoders extend standard text decoders' capabilities by simultaneously processing visual and textual inputs. They are designed to handle the complexities of integrating information from multiple modalities. Role in the Framework 1. Multimodal text decoders create joint representations that combine visual and textual features. This holistic representation captures the intricate relationships between the two modalities. 2. These decoders utilize the attention mechanism to focus on the most relevant parts of the image and the text. This ensures that the model pays equal attention to critical visual details and essential textual information. 3. By integrating visual and textual data, multimodal text decoders enhance the model's contextual understanding, allowing it to make more informed decisions during classification and detection tasks. Model Training - Self-Supervised Learning: The framework employs self-supervised learning techniques, where the model learns from the data without requiring manual annotations for every feature. - Fine-Tuning: The model is fine-tuned using labelled data to improve its accuracy and performance. Results The new method was evaluated against standard benchmarks for insect-related tasks: - The proposed model achieved state-of-the-art performance, outperforming existing methods. - The model significantly improved in capturing fine details and accuracy. Final Thoughts This paper introduces the Insect-1M dataset and a novel Insect Foundation Model. The Insect-1M dataset, with 1 million images across 34,212 species, includes detailed hierarchical taxonomic labels and descriptions, addressing the limitations of existing datasets in size and diversity. The Insect Foundation Model utilizes Patch-wise Relevant Attention to focus on critical image regions and Description Consistency Loss to align visual features with textual descriptions. These techniques significantly improve fine-grained insect classification and detection. Overall, the contributions of the Insect-1M dataset and the Insect Foundation Model advance the state-of-the-art in visual recognition, enhancing accuracy and detail capture. You can learn more here: Paper Project page If you're going to be at CVPR this year, be sure to come and say "Hi!"

https://medium.com/voxel51/patch-wise-attention-enhances-fine-grained-visual-recognition-6f87550b590e

datascienceharp

Lukas Höllein on the Challenges and Opportunities of Text-to-3D with “ViewDiff”

A Q&A with an author of a CVPR 2024 paper discussing the implications of his work for 3D Modeling

Lukas Höllein on the Challenges and Opportunities of Text-to-3D with "ViewDiff" A Q&A with an author of a CVPR 2024 paper discussing the implications of his work for 3D Modeling I got a chance to have a (virtual) sit-down Q&A session with Lukas Höllein about his paper ViewDiff: 3D-Consistent Image Generation with Text-to-Image Models, one of the accepted forpapers CVPR 2024. His paper introduces ViewDiff, a method that leverages pretrained text-to-image models to generate high-quality, multi-view consistent images of 3D objects in realistic surroundings by integrating 3D volume-rendering and cross-frame-attention layers into a U-Net architecture. Lukas discusses the challenges of training 3D models, the innovative integration of 3D components into a U-Net architecture, and the potential for democratizing 3D content creation. Hope you enjoy it! Harpreet: Could you briefly overview your paper's central hypothesis and the problem it addresses? How does this problem impact the broader field of deep learning? Lukas: Pretrained text-to-image models are powerful because they are trained on billions of text-image pairs. In contrast, 3D deep learning is largely bottlenecked by much smaller datasets. Training models on 3D datasets will reach a different quality and diversity than we have nowadays in 2D. This paper shows how to bridge this gap: we take a model trained on 2D data and only finetune it on 3D data. This allows us to keep around the expressiveness of the existing model but translate it into 3D. Harpreet: Your paper introduces a method that leverages pretrained text-to-image models as a prior, integrating 3D volume-rendering and cross-frame-attention layers into each block of the existing U-Net network. What are the key innovations of this technique, and how does it improve upon existing methods? Lukas: The key innovation shows how we can utilize the text-to-image model and still produce multi-view consistent images. Earlier 3D generative methods created some 3D representations and rendered images from them. Integrating a text-to-image model into this pipeline is problematic because it operates on different modalities (images vs. 3D). In contrast, we keep around the 2D U-Net architecture and only add 3D components. By design, this allows the creation of consistent 3D images. Our output is *not* a 3D representation but multi-view consistent images (that can be turned into such a representation later). Harpreet: One of the significant findings in your research is the ability to generate multi-view consistent images that are photorealistic and diverse. Can you explain the implications of this result for real-world applications in deep learning? Lukas: Eventually, we want to be able to create entire 3D scenes with the help of pretrained deep learning models. This would significantly reduce the time and skills required (e.g. instead of hiring expert artists in 3D modelling). Basically, it democratizes 3D content creation. One example I like is sending GIFs to friends through messengers. How cool would it be to create your own just from text input? Our paper is one step in that direction. By specifying a text prompt, people can use such methods to create 3D assets and their corresponding surroundings. Harpreet: What challenges did you face during your research, particularly in implementing or validating the integration of 3D volume-rendering and cross-frame-attention layers into the U-Net architecture? How did you overcome them? Lukas: Issue 1: Make images consistent → It turns out that both 3D volume rendering and cross-frame attention are necessary. The first gives accurate control over poses. Without it, the generated images do not closely follow the input poses. The second ensures a consistent object identity. Issue 2: Keep around 2D prior → We want text prompt control, but we finetuned on a smaller 3D dataset. We used the Dreambooth paper's trick to finetune a prior preservation dataset. Harpreet: For practitioners looking to apply your findings, what practical steps or considerations should they consider? Are there specific scenarios where your method shines the most? Lukas: Our method needs a lot of memory to be trained, but it can run at inference time on smaller GPUs. Remember the desired output domain: a single category of objects or generalized across a dataset of multiple categories. This influences the training time. Limitations: flickering due to lighting differences → can reduce it with better data. Harpreet: The quality and diversity of training data are crucial for the effectiveness of diffusion models. Can you discuss your approach to data collection, cleaning, and curation to ensure the data is well-prepared and representative? How do you address challenges regarding ensuring fairness and minimizing bias in your datasets? Lukas: 1. Data Collection and Cleaning: Real-world Video Capture: We capture real-world videos of diverse objects and scenes. This provides a rich source of data that reflects the complexity of the real world. Image Extraction and Filtering: We extract individual frames from the videos and employ a filtering process to ensure high quality and remove blurry or otherwise unusable frames. This step is essential for creating a clean and reliable dataset. 2. Data Curation for Specific Control Mechanisms: 3D Pose Control: We aim to enable control over the 3D pose of generated objects. To achieve this, we align videos of different objects into a shared world space. This allows us to consistently manipulate objects' pose within the model's training data. Text-based Control: We want to enable users to control the generated output through text prompts. To facilitate this, we label images with a pre-trained image captioning model. This provides a textual representation of the image content, which can be used for text-based control. To further ensure diversity in the output, we generate multiple captions per image and sample them randomly during training. 3. Mitigating Bias: Pose Control Fairness: A key challenge is ensuring fairness in our pose control mechanism. We aim to avoid biases where certain poses are overrepresented in the training data. To address this, we implement a sampling strategy that ensures every pose direction is sampled equally often. This helps to prevent the model from learning biased representations of object poses. Final Thoughts This Q&A with Lukas Höllein, author of the CVPR 2024 paper "ViewDiff," highlights the potential of leveraging pretrained text-to-image models for 3D generation. ViewDiff's approach, integrating 3D components into a U-Net architecture, addresses the challenges of training 3D models and demonstrates the feasibility of generating multi-view consistent images from text prompts. The method's ability to generate realistic 3D scenes and assets has significant implications for democratizing 3D content creation. ViewDiff represents a significant advancement in the field, paving the way for further research and development in text-to-3D generation. You can learn more about ViewDiff here: Paper GitHub Project Page If you'll be at CVPR this year, come and say "Hi!"

https://medium.com/voxel51/lukas-h%C3%B6llein-on-the-challenges-and-opportunities-of-text-to-3d-with-viewdiff-40203fb59c93

datascienceharp

Fixing CLIP’s Blind Spots: How New Research Tackles AI’s Visual Misinterpretations

A CVPR Paper Review and Cliff’s Notes

Fixing CLIP's Blind Spots: How New Research Tackles AI's Visual Misinterpretations A CVPR Paper Review and Cliff's Notes Overview The paper "Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs" investigates the visual question-answering (VQA) capabilities of advanced multimodal large language models (MLLMs), particularly focusing on GPT-4V. It highlights systematic shortcomings in these models' visual understanding and proposes a benchmark for evaluating their performance. The authors introduce the Multimodal Visual Patterns (MMVP) benchmark and propose a Mixture of Features (MoF) approach to improve visual grounding in MLLMs. No time to read the blog? No worries! Here's a video of me covering what's in this blog! Existing Challenge Despite their impressive capabilities, multimodal AI models like GPT-4V often fail to correctly answer basic questions about images. These failures are mostly due to how these models interpret visual information. Why Current Methods Fail The current methods rely heavily on a system called CLIP. CLIP pairs images with text descriptions to create a joint understanding of both. However, CLIP has a significant flaw: it can create what's known as "CLIP-blind pairs." CLIP-Blind Pairs Definition: CLIP-blind pairs are sets of images that CLIP sees as very similar, even though they are quite different. Example: Imagine two images, one of a cat and one of a dog. If CLIP considers these images similar because they both have furry animals, it might treat them as nearly identical, even though cats and dogs are very different. Impact: This confusion leads to poor visual representations. When the multimodal model tries to answer questions about these images, it might confuse details or provide incorrect answers because it doesn't truly understand the visual differences. These issues with CLIP-blind pairs propagate to more advanced models that use CLIP as their visual backbone. As a result, these models: Give Incorrect Answers: They might misidentify objects or misunderstand their positions in the image. Hallucinate Explanations: They sometimes make up explanations for their incorrect answers, which can be misleading. The Solution: Mixture of Features (MoF) Proposed Solution The researchers introduced the Mixture of Features (MoF) approach to tackle these visual shortcomings. MoF aims to improve the visual grounding capabilities of these models by integrating better visual representations. How the Solution Works Current Method (CLIP): CLIP tries to understand images by comparing them to text descriptions, but it struggles with CLIP-blind pairs, leading to ambiguous or incorrect visual representations. Improvements with MoF: Additive-MoF (A-MoF): This method combines features from CLIP with another system called DINOv2. The model's overall visual grounding improves by adding features from DINOv2, which better understand visual details. However, this can sometimes reduce the model's ability to follow text instructions precisely. Interleaved-MoF (I-MoF): This method spatially mixes visual tokens from CLIP and DINOv2. This more integrated approach ensures that the model benefits from the detailed visual understanding of DINOv2 while maintaining its capability to follow instructions from text. Why It's Better The MoF approach offers several benefits: Improved Visual Understanding: By incorporating features from DINOv2, the models become better at distinguishing details in images, reducing errors from CLIP-blind pairs. Balanced Capabilities: The Interleaved-MoF method ensures that the models understand images and follow text instructions. Systematic Error Reduction: This approach directly addresses the visual confusion caused by CLIP-blind pairs, leading to more accurate answers. Key Contributions The main contributions of the paper include: Detailed Analysis: An in-depth study of the visual shortcomings in current multimodal models, particularly those based on CLIP. New Testing Tool: The Multimodal Visual Patterns (MMVP) benchmark has been introduced to better evaluate how well these models understand images. Improved Method: The development of the Mixture of Features (MoF) approach, which integrates different types of visual understanding to enhance model performance. Results The researchers tested their new method and found: All the models, including GPT-4V, struggled with simple visual questions. GPT-4V performed better than random guessing but still had significant room for improvement compared to humans. The MoF approach significantly improved visual grounding, reducing errors caused by CLIP-blind pairs. Real-World Applications A better visual understanding of AI models can be useful in many fields: Animation and Gaming: It can help create more realistic characters and interactions. Virtual and Augmented Reality: It can make VR/AR environments more accurate and immersive. Retail and Online Shopping: It can improve product searches and recommendations. Final Thoughts The improvements suggested in the paper are important because they improve AI models' understanding of images, which is crucial for many applications. This research helps make high-quality visual understanding more accessible and reliable. Learn more about the paper by visiting: Project Page GitHub If you're going to be at CVPR this year, be sure to come and say "Hi!"

https://medium.com/voxel51/fixing-clips-blind-spots-how-new-research-tackles-ai-s-visual-misinterpretations-8b8ef4b4c250

datascienceharp

Interval Score Matching: Enhancing Fidelity in Text-to-3D Models with LucidDreamer

A CVPR Paper Review and Cliff’s Notes

Interval Score Matching: Enhancing Fidelity in Text-to-3D Models with LucidDreamer A CVPR Paper Review and Cliff's Notes Traditional 3D modelling is time-consuming and requires specialized skills, creating a barrier to widespread use in various industries. Recent advancements in text-to-3D generation have shown promise yet often fail to produce models with fine details and realism. Addressing these challenges, the latest research introduces novel methodologies to bridge this gap. This paper introduces LucidDreamer, a new system that can create detailed and realistic 3D models from text descriptions. Imagine you could type "a red sports car" into a system and, within minutes, receive a highly detailed 3D model that captures the intricate curves, reflective surfaces, and precise proportions of a real sports car. No time to read the blog? You can hear me talk about the paper: LucidDreamer, with its Interval Score Matching (ISM) technique, achieves this level of high-fidelity text-to-3D generation. By addressing the limitations of previous methods like Score Distillation Sampling (SDS), LucidDreamer produces 3D models with unparalleled detail and realism, making it a groundbreaking tool for applications ranging from virtual reality to digital content creation. The Problem Creating 3D models is usually a time-consuming task that requires expertise. Several advancements have recently allowed us to generate 3D models from text descriptions, for example: Magic3D A text-to-3D content creation tool developed by NVIDIA that generates high-quality 3D mesh models from textual descriptions. It utilizes image conditioning techniques and a prompt-based editing approach to provide users with novel ways to control 3D synthesis. Fantasia3D A text-to-3D content creation that disentangles geometry and appearance modelling, enabling the generation of photorealistic 3D assets from text prompts. It uses a hybrid scene representation and encodes surface normals extracted from the representation as input to an image diffusion model for geometry learning. ProlificDreamer A text-to-3D generation method that uses variational score distillation to generate high-fidelity and diverse 3D content from text prompts. It improves upon the existing score distillation sampling (SDS) method by modelling the 3D parameter as a random variable instead of a constant, addressing issues like over-saturation, over-smoothing, and low diversity in generated 3D models. Still, these methods often produce models that are not very detailed or realistic. One popular method for this is called Score Distillation Sampling (SDS), but it has some issues: The models it creates can look "smooth" and lack detail. The updates it makes to improve the 3D model are often inconsistent. The Solution: Interval Score Matching (ISM) To solve these problems, the authors propose a new approach called Interval Score Matching (ISM). Let's break down how this works: Score Distillation Sampling (SDS): First, it's essential to understand that SDS uses a pre-trained model that can convert text to images. It tries to use this model to guide the creation of a 3D model. However, the way it updates the 3D model tends to average out details, making the final result look smooth and not very detailed. ISM Improvements: DDIM Inversion: This is a fancy way of saying that ISM uses a method to create a consistent path for updating the 3D model, reducing randomness and improving detail. Interval-Based Matching: Instead of making big jumps in updating the 3D model, ISM makes smaller, more controlled updates. This helps maintain the details and avoid errors. Why It's Better With these improvements, LucidDreamer can create 3D models that are much more detailed and realistic compared to older methods. It also does this more efficiently, requiring less time and computing power. Key Contributions Detailed Analysis: The authors examined why SDS wasn't working well and identified its fundamental problems. New Method (ISM): They introduced ISM, which significantly improves the quality of 3D models. Advanced Techniques: By combining ISM with 3D Gaussian Splatting, they enhanced the 3D model quality by reducing the training time. Results The new method (LucidDreamer using ISM) was tested and shown to produce better and more detailed 3D models compared to other state-of-the-art methods like Magic3D, Fantasia3D, and ProlificDreamer. Plus, it does this with less training, making it more efficient. Real-World Applications This technology can be used in various fields, including: Animation and Gaming: Creating detailed characters and environments. Virtual and Augmented Reality: Building realistic 3D assets for VR and AR experiences. Retail and Online Shopping: Generating 3D models of products based on descriptions. Final Thoughts The paper introduces significant improvements in generating 3D models from text, making the process faster and producing better-quality results. This makes it easier for people without 3D modelling skills to create high-quality 3D content. Learn More The authors mentioned they will make their code available online, meaning others can use and build upon it. This is great for the research community and developers interested in this technology. GitHub If you're going to be at CVPR this year, be sure to come and say "Hi!"

https://medium.com/voxel51/interval-score-matching-enhancing-fidelity-in-text-to-3d-models-with-luciddreamer-f18c022dd4ac

datascienceharp