---
annotations_creators: []
language: en
license: bsd-3-clause
size_categories:
- 1K<n<10K
task_categories:
- object-detection
- image-to-text
task_ids: []
pretty_name: Total-Text-Dataset
tags:
- fiftyone
- image
- object-detection
- text-detection
dataset_summary: >



  ![image/png](dataset_preview.jpg)



  This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 1555
  samples.


  ## Installation


  If you haven't already, install FiftyOne:


  ```bash

  pip install -U fiftyone

  ```


  ## Usage


  ```python

  import fiftyone as fo

  import fiftyone.utils.huggingface as fouh


  # Load the dataset

  # Note: other available arguments include 'max_samples', etc

  dataset = fouh.load_from_hub("Voxel51/Total-Text-Dataset")


  # Launch the App

  session = fo.launch_app(dataset)

  ```
---

# Dataset Card for Total-Text-Dataset

The Total-Text consists of 1555 images with more than 3 different text orientations: Horizontal, Multi-Oriented, and Curved


![image/png](dataset_preview.jpg)


This is a [FiftyOne](https://github.com/voxel51/fiftyone) dataset with 1555 samples.

## Installation

If you haven't already, install FiftyOne:

```bash
pip install -U fiftyone
```

## Usage

```python
import fiftyone as fo
import fiftyone.utils.huggingface as fouh

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = fouh.load_from_hub("Voxel51/Total-Text-Dataset")

# Launch the App
session = fo.launch_app(dataset)
```

## Dataset Details

### Dataset Description

<!-- Provide a longer summary of what this dataset is. -->

- **Curated by :** Chee-Kheng Ch’ng, Chee Seng Chan, Cheng-Lin Liu
- **Funded by :** Fundamental Research Grant Scheme (FRGS) MoHE (Grant No. FP004-2016) and Postgraduate Research Grant (PPP) (Grant No. PG350-2016A).
- **Language(s) (NLP):** en
- **License:** bsd-3-clause

### Dataset Sources

<!-- Provide the basic links for the dataset. -->

- **Repository :** https://github.com/cs-chan/Total-Text-Dataset
- **Paper :** https://arxiv.org/abs/1710.10400

## Uses

- curved text detection problems

## Dataset Structure

```
Name:        Total-Text-Dataset
Media type:  image
Num samples: 1555
Persistent:  False
Tags:        []
Sample fields:
    id:                     fiftyone.core.fields.ObjectIdField
    filepath:               fiftyone.core.fields.StringField
    tags:                   fiftyone.core.fields.ListField(fiftyone.core.fields.StringField)
    metadata:               fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.metadata.ImageMetadata)
    ground_truth_polylines: fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.labels.Polylines)
    ground_truth:           fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.labels.Detections)
```
The dataset has 2 splits: "Train" and "Test". Samples are tagged with their split.

## Dataset Creation

### Curation Rationale

At present, text orientation is not diverse enough in the existing scene text datasets. Specifically, curve-orientated text is
largely out-numbered by horizontal and multi-oriented text, hence, it has received minimal attention from the community so
far. Motivated by this phenomenon, the authors collected a new scene text dataset, Total-Text, which emphasized on text orientations
diversity. It is the first relatively large scale scene text dataset that features three different text orientations: horizontal, multioriented,
and curve-oriented.

#### Annotation process

<!-- This section describes the annotation process such as annotation tools used in the process, the amount of data annotated, annotation guidelines provided to the annotators, interannotator statistics, annotation validation, etc. -->
Initial version of Total-Text’s polygon annotation was carried out with the mindset of covering text instances tightly with
the least amount of vertices. As a result, the uncontrolled length of polygon vertices is not practical to train a regression
network. The authors refined the Total-Text annotation using the following scheme. Apart from setting the number 
of polygon vertices to 10 (empirically, 10 vertices are found to be sufficient in covering all the word-level text instances
tightly in our dataset), they used a guidance concept inspired by Curved scene text detection via transverse and longitudinal 
sequence connection paper by Liu, et al. which was introduced to remove human annotators’ bias and in turn producing a more consistent ground truth.
The process for other annotations can be referred from paper. 

The authors have mentioned in the paper that the human annotator was given the freedom to take a break whenever he feels like to, 
ensuring that he will not suffer from fatigue which in turn introduces bias to the experiment. Both time and annotation quality 
were measured internally (within the script) and individually to each image.

The authors have also proposed aided scene text detection annotation tool, T3, could help in providing a better scene text dataset in terms of quality and scale. 

#### Who are the annotators?

<!-- This section describes the people or systems who created the annotations. -->

Chee-Kheng Ch’ng, Chee Seng Chan, Cheng-Lin Liu and Chun Chet Ng 

## Citation

<!-- If there is a paper or blog post introducing the dataset, the APA and Bibtex information for that should go in this section. -->

**BibTeX:**

```bibtex
@article{CK2019,
  author    = {Chee Kheng Ch’ng and
               Chee Seng Chan and
               Chenglin Liu},
  title     = {Total-Text: Towards Orientation Robustness in Scene Text Detection},
  journal   = {International Journal on Document Analysis and Recognition (IJDAR)},
  volume    = {23},
  pages     = {31-52},
  year      = {2020},
  doi       = {10.1007/s10032-019-00334-z},
}
```

## Dataset Card Authors

[Kishan Savant](https://huggingface.co/NeoKish)