HumanMM: Global Human Motion Recovery from Multi-shot Videos

github.io/HumanMM Figure 1 . Recovering a human motion from multi-shot videos. Top: We take two multi-shot table tennis game videos with shot transitions as input. We aim to recover two motions of two athletes (Long MA and Zhendong FAN) from two videos, respectively. The first video is recorded by three shots ("x", "y", and "z" ), and the second one is recovered by two shots ("" and "|" ). Bottom: We recover two motions (Long MA in green and Zhendong FAN in pink), different shots, and camera poses for each multi-shot video. The recovered motion is aligned with the motion in the videos.