EgoActionBench: Benchmarking Dense Action Captioning for Egocentric Human Videos

Introduction

EgoActionBench is a benchmark for dense action captioning in egocentric human videos. It represents continuous human manipulation as temporally localized atomic actions, each associated with the left hand, right hand, or both hands. Given a clip, a model must recover what the camera wearer does, when each action happens, and which hand performs it.

EgoActionBench contains 1,407 episodes with 41,132 human-refined action annotations.

Release Status

Benchmark annotations and the episode index are coming soon. Visit our project page for annotation examples and benchmark results, and our GitHub repository for release updates.

Evaluation

EgoActionBench evaluates semantic fidelity, temporal localization, and hand-action association. Its progressive evaluation protocol retrieves temporally eligible predictions and selects a prediction for each reference action by content agreement. Semantic fidelity and temporal overlap are scored alongside action precision and recall, summarized as EgoActionScore.

See the evaluation protocol overview for details.

Data Sources

EgoActionBench builds on videos from Ego4D, EgoLife, and EPIC-KITCHENS-100. Access to the source videos is governed by their original licenses and terms of use.