Temporal Decoupling Decoding and Inverse Processing for Human Motion Prediction
How separating reconstruction and prediction into distinct decoders, rather than coupling them in one, resolves task conflict in skeleton-based human motion prediction. Students learn why a shared decoder must balance the past-behavior and future-behavior manifolds, and how temporally reversing and reintroducing motion (inverse processing) exploits bidirectional temporal correlation to sharpen predictions across model backbones.
2501.00315
TD2IP is a method for skeleton-based human motion prediction that separates the auxiliary reconstruction (historical) and the primary prediction (future) decoding into two distinct decoders sharing a…