Xiaowei Zhou

Reducing inter-view redundancy, PointSplat achieves superior 3D human representation quality while minimizing computational load.

Monocular human motion capture can be dramatically improved by explicitly modeling high-order temporal dynamics like velocity and acceleration, leading to more realistic and less jittery movements.

Robots can now learn flexible, geometry-free interactions with objects directly from video, sidestepping the need for laborious 3D modeling or complex retargeting.

Synthesizing realistic duet dance motions gets a boost from explicitly modeling inter-dancer contact, leading to significantly improved interaction fidelity and rhythmic synchronization.