Over 3,700 hours of first-person video, wrapped in Python

What it does The repository is a Python convenience layer around two Meta video datasets: Ego4D (annotated first-person footage totaling over 3,700 hours) and Ego-Exo4D (time-synchronized egocentric Aria and exocentric GoPro views, including 3D). It bundles CLI downloaders, a unified video-reading API atop TorchAudio and PyAV, feature-extraction wrappers for models like Omnivore and SlowFast, and Jupyter notebooks to jump-start analysis.

The interesting bit The value is in the plumbing. Meta has turned two large-scale, multi-modal video corpora into an importable module with pre-built model wrappers and a visualization engine, so you spend less time fighting AV codecs and more time training.

Key highlights

• Downloaders for both Ego4D and Ego-Exo4D (the latter now at 1,286.30 video hours, 221.26 of them ego-centric, in V2).
• A single video-I/O API abstracting TorchAudio and PyAV.
• Lightweight feature-extraction wrappers for off-the-shelf models including Omnivore and SlowFast.
• Jupyter notebooks and a web-based visualizer for dataset exploration.
• Research code such as CLEP (Contrastive Language Ego-centric video Pre-training).

Caveats

• Baseline code for Ego-Exo4D benchmarks is marked “coming soon,” so don’t expect ready-made challenge starters here yet.
• Ego4D benchmark baselines live in separate repositories under the EGO4D GitHub organization, not in this repo.
• The visualizer requires a dataset license to use, even though the repository itself is MIT-licensed.

Verdict Grab it if you’re doing egocentric or multi-view video research and want to skip the usual data-acquisition slog. Look elsewhere if you need a fully self-contained benchmark suite—much of the baseline code is scattered across other repos.