Ego-Exo4D Proficiency

Classified by the Atlas from how the authors describe and distribute it.

Crossref lists 101 authors, first author Kristen Grauman (https://api.crossref.org/works/10.1007/s11263-025-02557-6).

Release date from https://docs.ego-exo4d-data.org/changelog/.

Commit messages: 'test split video_ids', 'updated v2 test set videos'. IJCV date from https://api.crossref.org/works/10.1007/s11263-025-02557-6.

Classification and temporal localisation on recorded video.

Basketball, bike repair, cooking, dance, health (COVID testing), music, rock climbing, soccer.

Demonstrator proficiency: 1,983 train / 561 val / 768 test takes, 6 scenarios (bike repair and health excluded), 4 classes (novice, early, intermediate, late expert). Demonstration proficiency: 556 train / 177 val / 179 test takes, 8 scenarios.¹

Parent dataset: 740 participants, 13 cities, 123 scene contexts, 1,286 hours, 5,035 takes (arXiv v4 and V2 data). arXiv v1 said >800 participants, 131 scene contexts, 1,422 hours.¹

Docs list mAP at k = 0.25, 0.5, 1.0 s (https://docs.ego-exo4d-data.org/benchmarks/proficiency_estimation/). Challenge numbers are each team's own report (https://arxiv.org/abs/2505.24411, https://arxiv.org/abs/2507.08022); official leaderboard page not found.

A Proficiency Estimation Challenge ran at the CVPR 2025 EgoVis workshop (participant reports), but no leaderboard page was found. The 2026 challenge page lists only EgoPose Body and Procedure Understanding (CodaBench) and says organisers are moving from EvalAI to CodaBench.

EGO4D/ego-exo4d-proficiency has no LICENSE file (GitHub API license null). The dataset download tool repo facebookresearch/Ego4d is MIT (https://github.com/facebookresearch/Ego4d).

Signed per individual or institution at ego4d.dev (request page returned HTTP 403 to us). Docs say it covers 'research purposes and commercial use' with restrictions on redistribution; full terms not read.

Approval typically ~48 hours; AWS credentials expire after 14 days.

Crossref metadata (DOI 10.1007/s11263-025-02557-6), article licence CC BY-NC-ND 4.0. CVPR 2024 stated in arXiv comments (https://arxiv.org/abs/2311.18259). Springer and CVF pages blocked automated access.