HiERO-StepG @ Ego4D Step Grounding Challenge: hierarchical activity understanding enables zero-shot step grounding

Researchers introduced HiERO-StepG, a zero-shot method for Ego4D step grounding built on weakly supervised hierarchical action representations.

The system uses clustering to detect procedure steps, plus constraints and post-processing to refine temporal predictions.

It reached 56.27% R@1 at IoU 0.3 on the leaderboard and ranked second at submission time.

The result suggests step grounding in procedural videos can work without procedure-specific labels, lowering annotation costs.