AbhayARao26/FinalSignsB4Aug-ISL

Dataset Summary

This dataset contains approximately 1,300 videos organized into 138 distinct categories/folders. Each folder contains about 9 to 10 videos representing a specific sign, action, or concept. The total dataset size is approximately 2.5 GB.

The folder names are the class labels. Most labels are written in Devanagari script, and some classes have duplicate or variant folders such as खाना1, बस1, or दर्द1.

Supported Tasks and Leaderboards

• Video Classification: Train models to classify video clips into 138 distinct classes.

Dataset Structure

The dataset is organized hierarchically, where each top-level folder corresponds to one class label.

dataset_root/
├── अच्छा/
│   ├── sample_1.mp4
│   └── sample_2.mp4
├── अंदर/
│   ├── sample_1.mp4
│   └── sample_2.mp4
├── ...
└── हाथ/
    ├── sample_1.mp4
    └── sample_2.mp4

Data Instances

Each instance is a short video clip belonging to one sign class. The dataset currently contains:

• 138 class folders
• Approximately 1,262 .mp4 files
• 3 .MOV files

Data Fields

The dataset does not include a separate annotation file. The label for each video is inferred from its parent folder name.

• video: the video file itself
• label: the name of the enclosing folder

Data Splits

No predefined train/validation/test split is included. Users should create their own split, for example an 80/10/10 or 70/15/15 partition, while keeping class balance in mind.

Licensing Information

This dataset is released under the MIT License.

Limitations

• Class names are stored as folder names, so users must map folder labels to class IDs during preprocessing.
• The dataset contains a small number of non-MP4 files and system files such as .DS_Store; these should be filtered out before training.
• Because the dataset is relatively small for deep video learning, augmentation and careful validation are recommended.

Recommended Citation

If you use this dataset, please cite it as:

Final_ISL_Signs_138. 2026. Released under the MIT License.