Improving Viewpoint-Invariance and Temporal Consistency for Action Detection

Researchers introduced a two-stage human action detection framework for untrimmed videos.

The method uses training-time virtual viewpoints plus a view-invariant temporal encoder to handle camera-angle changes and long-range motion.

It leverages motion features and a multi-scale temporal encoder, including a selective state-space model, for stronger temporal consistency.

The approach reported strong benchmark gains on PKU-MMD and BABEL, outperforming prior work.

The results could improve real-world video understanding systems that must generalize across viewpoints and long sequences.