PoseBridge: Bridging the Skeletonization Gap for Zero-Shot Skeleton-Based Action Recognition
PoseBridge fixes a problem where zero-shot action recognition loses crucial object and pose information when converting human poses into skeletons. It creates a framework that uses intermediate human pose estimation features to extract 'pose-anchored' semantic cues, which are then used to bridge the gap between skeleton motion and text semantics, leading to better zero-shot performance.