DINOv2

Video dynamic signals are often tightly intertwined between camera displacement and object motion. Through self-supervised learning, SDM decomposes features into primary and residual tokens, improving motion-prediction accuracy in ProbeMotion evaluations.

影片動態訊號常在相機位移與物體運動間高度交織。SDM 透過自監督學習，將特徵拆解為主要與殘差 token，在 ProbeMotion 評估中提升了運動預測準確度。