Wanqing Li

Replacing deterministic coordinate regression with masked joint diffusion captures the subtle, long-tail dynamics standard skeleton encoders miss, driving a 1.1% gain on competitive action recognition benchmarks.

Adding more views can actually hurt performance—this research reveals how to intelligently fuse motion cues for superior action quality assessment.

Synthetic pose data can significantly boost skeleton-based action recognition, even rivaling performance with real data, thanks to a new conditional generative pipeline.