Velocity-to-velocity human motion forecasting

Hongsong Wang, Liang Wang, Jiashi Feng, Daquan Zhou

Abstract: Highlights•We introduce a novel velocity-to-velocity learning paradigm for human motion prediction, and propose different architectures to implement this paradigm.•We design an end-to-end trainable RMT layer which transforms joint angles from the exponential map to the 3D rotation matrix.•We define a novel robust loss function in the space of 3D rotation matrices.•We present a robust training algorithm which exploits several sequence transformation techniques such as Gaussian smoothing.