YijiaLeithu/EgoSAT-ROI-Cache

EgoSAT ROI Cache

Dataset Description

This dataset contains ROI cache files for EgoSAT. The cache is intended for ROI-aware streaming vision-language models such as ROI-TimeChat and related hand/object/gaze-aware experiments.

This dataset does not contain raw Ego4D videos. Users must obtain Ego4D access and license approval separately before running any workflow that reads the underlying videos.

Related resources:

• Paper: https://arxiv.org/abs/2606.24422
• Code: https://github.com/leiyj23/EgoSAT
• Project page: https://leiyj23.github.io/EgoSAT/
• Main dataset repo: https://huggingface.co/datasets/YijiaLeithu/EgoSAT
• ROI cache dataset repo: https://huggingface.co/datasets/YijiaLeithu/EgoSAT-ROI-Cache

DOI, proceedings volume, and page numbers will be added after proceedings publication.

Dataset Structure

.
  README.md
  roi_cache/
  metadata/
    roi_cache_index.jsonl
    layout_version.json
    checksums.json
    checksums_all.json
    release_manifest.json

• roi_cache/: per-sample cache files in JSONL format.
• metadata/roi_cache_index.jsonl: index mapping cache filenames to video and clip metadata.
• metadata/layout_version.json: filename pattern, coordinate convention, and cache layout version.
• metadata/checksums.json and metadata/checksums_all.json: file checksums.
• metadata/release_manifest.json: release-candidate inclusion and exclusion manifest.

Alignment

Cache filenames follow:

<video_uid>__<clip_id>__<clip_uid>.roi_cache_merged_fps1.jsonl

Use the filename components or metadata/roi_cache_index.jsonl to align a cache file with the corresponding EgoSAT sample. ROI cache alignment is based on video_uid, clip_id, clip_uid, and filename/index metadata.

ROI Record Schema

metadata/layout_version.json records:

• layout_version: roi_cache_merged_fps1.
• record_format: jsonl.
• coord: norm.
• filename_pattern: <video_uid>__<clip_id>__<clip_uid>.roi_cache_merged_fps1.jsonl.

When coord is norm, box and gaze coordinates are normalized to the frame coordinate system.

metadata/roi_cache_index.jsonl is a JSONL index. Sampled rows contain:

• video_uid: Ego4D video UID.
• clip_id: EgoSAT clip or interval id.
• clip_uid: Ego4D clip UID.
• filename: cache filename.
• cache_path_relative: path to the cache file relative to this dataset root.
• size_bytes: cache file size.
• num_records: number of frame records in the cache file.
• format: file format.
• layout_version: cache layout version.
• sha256: cache file checksum.

Each file under roi_cache/ is organized frame by frame as JSONL. Sampled records contain:

• video_uid, clip_id, clip_uid: video and clip identifiers.
• frame_idx: frame index for the cache record.
• ts_sec: timestamp relative to the cache clip.
• t_abs_sec: absolute timestamp in the source video timeline.
• clip_start_sec_abs, clip_end_sec_abs: absolute clip boundaries.
• coord: coordinate convention, typically norm.
• hand_boxes: detected hand boxes.
• hand_scores: confidence scores for hand_boxes.
• hand_labels: hand labels corresponding to hand_boxes.
• left_box, left_score, right_box, right_score: side-specific hand boxes and scores when available.
• orig_frame_idx_hand, t_hand_abs, clip_t_hand, src_hand: source metadata for hand detections.
• gaze: gaze record when available, including fields such as x, y, r, and conf.
• orig_frame_idx_gaze, ts_sec_gaze_rel, t_abs_gaze_sec, src_gaze: source metadata for gaze predictions.
• src: merged source label, for example hand plus gaze provenance.

Detections may be empty or unavailable for some frames. Loaders should handle null, empty arrays, or missing optional fields.

Ego4D Videos

Raw Ego4D videos are not redistributed in this ROI cache dataset. The cache does not grant Ego4D video access or redistribution rights. Users must separately obtain Ego4D access and comply with the Ego4D license and terms of use before using this cache with video frames.

License and Usage Terms

Dataset license: other.

The EgoSAT ROI cache is released for research use. Raw Ego4D videos are not redistributed in this dataset. Users must separately obtain Ego4D access and comply with the Ego4D license and terms of use. This dataset does not grant any rights to the original Ego4D videos.

The code repository is released separately under the MIT License. Users are responsible for ensuring that their use complies with the licenses and terms of Ego4D and EgoSAT.

Citation

If you use EgoSAT or this ROI cache, cite the benchmark paper:

@inproceedings{lei2026egosat,
  title={EgoSAT: A Comprehensive Benchmark of Egocentric Streaming Interaction Understanding},
  author={Lei, Yijia and Li, Jinzhao and Zhang, Yichi and Hua, Jiacheng and Li, Yin and Liu, Miao},
  booktitle={European Conference on Computer Vision},
  year={2026}
}

The arXiv preprint can be cited as:

@article{lei2026egosat_arxiv,
  title={EgoSAT: A Comprehensive Benchmark of Egocentric Streaming Interaction Understanding},
  author={Lei, Yijia and Li, Jinzhao and Zhang, Yichi and Hua, Jiacheng and Li, Yin and Liu, Miao},
  journal={arXiv preprint arXiv:2606.24422},
  year={2026}
}

Responsible Use

Use this cache only for research and evaluation with properly licensed Ego4D video access and compatible EgoSAT annotations. Do not redistribute raw Ego4D videos. Do not attempt to reconstruct identities or infer private attributes from source videos. Do not treat benchmark outputs as safety-critical decisions.