CorrCLIP: Pre-generated SAM2 Masks for Open-Vocabulary Semantic Segmentation

This repository contains the pre-generated SAM2 masks used to replicate the results presented in the paper CorrCLIP: Reconstructing Patch Correlations in CLIP for Open-Vocabulary Semantic Segmentation.

The paper addresses the challenge of open-vocabulary semantic segmentation by proposing CorrCLIP, which reconstructs patch correlations in CLIP. It leverages the Segment Anything Model (SAM) to define the scope of patch interactions, aiming to improve alignment and reduce inter-class correlations. These masks are a crucial component for reproducing the paper's findings and are specifically mentioned for use when setting mask_generator to None in the project's configuration.

For the official code, detailed installation instructions, evaluation scripts, and further information on CorrCLIP, please refer to the project's GitHub repository: https://github.com/zdk258/CorrCLIP

Sample Usage

A Gradio demo is available to perform open-vocabulary semantic segmentation on custom images with custom category names. You can run it locally by following the instructions in the GitHub repository:

python demo_gradio.py

Additionally, a Google Colab demo is provided for easy online experimentation: https://colab.research.google.com/github/zdk258/CorrCLIP/blob/master/corrclip_demo.ipynb

Citation

If you find CorrCLIP or these pre-generated SAM2 masks useful for your research, please consider citing the original paper:

@article{zhang2024corrclip,
  title={Corrclip: Reconstructing patch correlations in clip for open-vocabulary semantic segmentation},
  author={Zhang, Dengke and Liu, Fagui and Tang, Quan},
  journal={arXiv preprint arXiv:2411.10086},
  year={2024}
}