Weiyu Du, Oleh Rybkin, Lingzhi Zhang, Jianbo Shi
ECCV 2020 Workshop on Perception Through Structured Generative Models
(http://generativeperception.com/)
link to paper
link to video presentation
Motion reconstruction on AMASS test set, from left to right: ground truth, reconstruction with Sine Motion Encoding, reconstruction with Bezier Motion Encoding.
Point-to-point motion generation on AMASS test set (30 frames), from left to right: ground truth, prediction by Sine Motion Encoding, prediction by Bezier Motion Encoding, latent linear interpolation baseline.
Point-to-point motion generation on Human3.6M test set (sub-sampled 5 times, 30 frames), from left to right: ground truth, prediction by Sine Motion Encoding(6 curves), prediction by Sine Motion Encoding(3 curves), prediction by Bezier Motion Encoding, latent linear interpolation baseline.
Motion generation at different lengths and frame rates on Human3.6M test set using Sine Motion Encoding (3 curves), from left to right: ground truth (30 frames, fps=12), prediction (30 frames, fps=12), prediction (60 frames, fps=24).
