gaze_dataset_full — StreamGaze_v2 + EgoGazeVQA + HD-EPIC

A single repository containing three complementary benchmarks for evaluating multimodal LLMs on gaze-grounded egocentric video question answering:

Subfolder	Source	Held-out (val/test)	Questions
StreamGaze_v2/	egoexolearn, holoassist, egtea	egtea	8 MCQ tasks (4-opt) — gaze-conditioned past/present/future
EgoGazeVQA/	ego4d, egoexo, egtea	egtea	causal / spatial / temporal (5-opt)
HD-EPIC/	P01–P09	P09 (participant)	30 vqa-benchmark tasks (5-opt)

CombinedMergeDataset in the TrajGazeMerge pipeline takes the union of all three sources, training on the non-held-out items and evaluating on the held-out ones.

---

Repository layout

gaze_dataset_full/
├── README.md
├── restore.sh                        unified restore — extracts all 3 sources
├── StreamGaze_v2/                    (~575 GB packed forms)
│   ├── frames_shards/                tar shards of per-clip frames
│   ├── videos_tars/                  pre-packed video tars
│   ├── gaze/, hand/, interaction/    per-frame gaze/hand and interaction npz
│   ├── adapted/, qa/, metadata/      QA and per-task metadata
│   └── README.md
├── EgoGazeVQA/                       (~225 GB packed forms)
│   ├── metadata.csv                  1750 MCQ items
│   ├── ego4d_gaze.tar, ego4d_no_gaze.tar, ego4d_gaze_mapping.tar
│   ├── egoexo_gaze_part1/2.tar, egoexo_no_gaze_part1/2.tar, egoexo_gaze_mapping.tar
│   ├── egtea_gaze.tar, egtea_no_gaze.tar, egtea_gaze_mapping.tar
│   ├── ego4d_jsons/, egoexo_jsons/, egtea_jsons/
│   ├── {ego4d,egoexo,egtea}/hand_locations/
│   └── interaction/
└── HD-EPIC/                          (~9.7 GB packed forms)
    ├── frames_extracted_tars/        9 tars — frames_extracted_P0X.tar (224×224, 10 fps)
    ├── hand_locations.tar            per-clip hand JSON (StreamGaze schema)
    ├── interaction/                  per-clip interaction.npz
    └── vqa-benchmark/                30 task JSON files

The HD-EPIC subset includes only the files actually consumed by the training/eval pipeline (HDEpicDataset in TrajGazeMerge/data/): frames_extracted/, hand_locations/, interaction/, and the VQA task JSONs. Raw video/, Audio-HDF5/, SLAM-and-Gaze/, and Hands-Masks/ are not included — recover them from the upstream HD-EPIC release if needed.

The original HDEpicDataset expects the VQA JSON at /workspace/hd-epic-annotations/vqa-benchmark/; this release relocates that folder under HD-EPIC/vqa-benchmark/ to keep one root per source. Adjust VQA_BASE accordingly or ln -s HD-EPIC/vqa-benchmark /workspace/hd-epic-annotations/vqa-benchmark.

---

Quick start

huggingface-cli download Peanuttoad/gaze_dataset_full --repo-type dataset \
    --local-dir ./gaze_dataset_full
cd gaze_dataset_full && bash restore.sh

restore.sh handles all three sources in one pass. Pass --keep-shards to keep the packed tar files after extraction.

---

Provenance

• StreamGaze_v2/ — also published standalone at Peanuttoad/StreamGaze_v2.
• EgoGazeVQA/ — also published standalone at Peanuttoad/gaze_dataset (no interaction/; that subdir is only present in the combined releases).
• 2-source combined release (sg + eg, no HD-EPIC): Peanuttoad/StreamGaze_EgoGazeVQA.
• HD-EPIC source: HD-EPIC official site.

License: CC-BY-4.0 for the combined release. Underlying source data (Ego4D, EgoExo, EGTEA, EgoExoLearn, HoloAssist, HD-EPIC) retains its own licenses.