You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

EgoLife-Long (Track B of UltraStream-Bench, release b-1.0rc)

Status: private, internal. Not for public release until EgoLife-Stream's upstream license has been independently re-checked by a human (see "License and citation" below) — this card and repo exist for internal storage/sharing only.

What this is

Track B of UltraStream-Bench: a causal, reactive streaming QA benchmark built on top of EgoLife-Stream, itself derived from the EgoLife project. 53 egocentric runs (>=120 minutes each), 158.29 video hours, 477 human-authored questions. Every question has a t_ask: a method under evaluation may only see video in [0, t_ask], never beyond it — this is the same reactive/causal contract used across both UltraStream-Bench tracks (see scripts/build_json.py's docstring for how this maps back to release/v1.0-rc/track_b in the source repo).

Video is included (added 2026-11-11, per explicit request): all 53 merged runs are under videos/<run_id>.mp4 in this repo (308.93 GB total, byte-identical to the original merged files — verified file count and total size against the local source right after upload). videos_manifest.csv's hf_path column gives each run's path here. segments_manifest.json is still included for provenance (which original EgoLife-Stream source clips were concatenated, in what order, at what offsets, to produce each videos/<run_id>.mp4).

Format

Each data/{dev,test}_{4,5}opt.json is one JSON object:

{
  "benchmark": "UltraStream", "track": "b_egolife", "version": "b-1.0rc",
  "split": "test", "n_options": 5,
  "videos": [
    {"video_id": "...", "source": "egolife", "run_id": "...", "wearer": "A1",
     "day": "DAY1", "video_relpath": "runs/<run_id>.mp4", "duration_s": ...,
     "class": "egolife",
     "questions": [
       {"qa_id": "...", "t_ask": ..., "question": "...",
        "options": ["...", "...", "...", "...", "Not observed yet"],
        "answer_idx": 2, "question_type": "LAST", "mode": "near",
        "delta_s": ..., "q_source": "human", "stratum": "PRIMARY"}
     ]}
  ]
}

video_relpath is relative to wherever you place the merged run MP4s (runs/<run_id>.mp4 matching EgoLife-Stream's own layout).

Two variants: 4-opt and 5-opt

{split}_4opt.json keeps EgoLife-Stream's original four choices (A-D). {split}_5opt.json adds a fifth option, "Not observed yet" — but E is never the gold answer on Track B: every Track B question is answerable at t_ask. The 5-opt variant exists so a method is evaluated under the exact same 5-option interface as Track A (where E sometimes is gold, for genuinely unobserved-yet questions); on Track B, a method choosing E is simply wrong, and that under-commit rate is itself a diagnostic.

Key stats

See STATS.md (full tables + figs/) for the complete breakdown. Headline:

  • 53/92 question-bearing EgoLife-Stream runs kept (run length >= 120 min), 158.29/210.83 hours, 477/573 questions.
  • PRIMARY=371, SECONDARY=106.
  • 4-opt gold letter balance: no letter exceeds 35% (highest: C at 26.4%); "always pick the most common letter" baseline = 26.4% accuracy.
  • 0 questions with t_ask < 5 minutes.
  • 44 duplicate-evidence groups / 93 questions, and 1 near-duplicate question pair (reported as diagnostics only, nothing filtered).

Baselines

See results/baselines.md — copied verbatim from the source repo's reproduced tables (release/v1.0-rc/evaluation/), not re-scored here.

Reconstructing runs from original EgoLife-Stream clips

videos_manifest.csv: one row per run (run_id, wearer, day, duration_s, n_questions, split, n_segments, merged_video, hf_path — hf_path is where that run's video lives in this repo). segments_manifest.json: {run_id: [{source_path, merged_start_sec, merged_end_sec, wallclock_start_sec}, ...]}, in merge order — concatenate source_path clips from the original EgoLife-Stream identity/day directories at those offsets to reproduce runs/<run_id>.mp4 exactly as built by EgoLife-Stream's own run_mapping_30min.relative.json (gap threshold 1800s, gaps collapsed). This information was available in the upstream dataset, so nothing here is marked as missing.

Use with ultrastream-bench

git clone <ultrastream-bench repo>
cd ultrastream-bench
# drop the merged run_id.mp4 files this card describes into runs/ under
# the directory configs/paths.yaml calls VIDEO_ROOT_B, then:
python scripts/validate.py
python scripts/run.py --method chance --track b --split dev

scripts/build_json.py (included here) is the exact script that produced these four JSON files from release/v0.3.2 + release/v1.0-rc/track_b in the source repo — re-run it there if the release is revised.

License and citation

This dataset's QUESTIONS are human-authored for EgoLife-Stream, which is itself derived from EgoLife. EgoLife-Stream's own dataset card (local copy: release/v1.0-rc/DATA_CARD.md in the source repo, and EgoLife-Stream/docs/DATASET_CARD.md in the raw dataset) states its questions are aligned from Ego-R1-Data's manual-2.9K QA set onto EgoLife clips. EgoLife's own top-level dataset card declares license: mit (read directly from its local README.md YAML frontmatter at /hps/1ecb8c84-b37d-4cc1-9d56-68db808c0018/duyntc2/datasets/EgoLife/README.md).

This repo is kept private and internal specifically because that MIT declaration has not yet been independently re-verified against EgoLife's official, canonical source (project page / GitHub / HF repo) by a human on this team — do not make it public until that check happens.

EgoLife citation:

@article{yang2025egolife,
  title={EgoLife: Towards Egocentric Life Assistant},
  author={Yang, Jingkang and Liu, Shuai and Guo, Hongming and Dong, Yuhao and
          Zhang, Xiamengwei and Zhang, Sicheng and Wang, Pengyun and Zhou, Zitang and
          Xie, Binzhu and Wang, Ziyue and Ouyang, Bei and Lin, Zhengyu and
          Cominelli, Marco and Cai, Zhongang and Zhang, Yuanhan and Zhang, Peiyuan and
          Hong, Fangzhou and Widmer, Joerg and Gringoli, Francesco and Yang, Lei and
          Li, Bo and Liu, Ziwei},
  journal={arXiv preprint arXiv:2503.03803},
  year={2025}
}

Project page: https://egolife-ai.github.io/ · Paper: https://arxiv.org/abs/2503.03803 · Code: https://github.com/egolife-ai/EgoLife

QA provenance note: questions are aligned from Ego-R1-Data's manual-2.9K onto EgoLife-Stream clips; no separately verified citation for Ego-R1-Data was found locally, so it is named here in prose only, not as a BibTeX entry, to avoid asserting a citation we have not checked.

Downloads last month
1

Paper for duyngtr16061999/EgoLife-Long-Stream