Structure

The structure of EgoMask (Benchmark) is as follows:

├─ egomask
    ├─ JPEGImages
    │   ├─ refego
    │   └─ egotracks
    ├─ annotations
    ├─ subset
    │   ├─ long
    │   │   ├─ meta_expressions.json
    │   │   └─ meta.json
    │   ├─ medium
    │   │   └─ meta_expressions.json
    │   └─ short
    │       ├─ meta_expressions.json
    │       └─ meta.json
    ├─ meta_expressions.json
    └─ meta.json

The structure of EgoMask-Train (Training dataset) is as follows:

├─ egomask-train
    ├─ meta_expressions
    │   └─ train
    │       └─ meta_expressions.json 
    └─ train
        ├─ JPEGImages
        ├─ mask_dict.json
        └─ meta.json

Dataset Prepare

Download Annotations

We have provided the annotations in Hugging Face.

hf download XuuuXYZ/EgoMask --repo-type dataset # --local-dir [PATH_TO_EGOMASK_DATA]

Process JPEGImages

Please follow Start here for instructions on how to access Ego4D dataset. Then use the following scripts from our repo to process images. Note that our aws profile name is set to ego4d.

For EgoMask (Benchmark)

bash preprocess/process_refego.sh
bash preprocess/process_egotracks_for_benchmark.sh

For EgoMask-Train (Training data)

bash preprocess/process_egotracks_fro_training.sh

🚩 Citation

If you find our EgoMask useful for your research, please consider giving this repository a star and citing our paper as follows:

@article{liang2025finegrained,
      title={Fine-grained Spatiotemporal Grounding on Egocentric Videos}, 
      author={Shuo Liang and Yiwu Zhong and Zi-Yuan Hu and Yeyao Tao and Liwei Wang},
      journal={arxiv preprint arXiv:2508.00518},
      year={2025},
}