DS-Seq: Deriving Smooth 3D Human Motion Sequences from Video Time Cues

Tao Peng, Delang Peng, Li Li, Junping Liu, Zili Zhang, Xinrong Hu

Abstract: When applying the methods for reconstructing the 3D poses and shapes of human bodies based on single-frame images to videos, the insufficient reconstruction accuracy of individual frames and the unbalanced utilization of information from the current frame, past frames, and future frames often lead to the inability to restore high-precision and smooth 3D human motion. Specifically, this results in irregular jitter in the generated human motion. To address this issue, we propose a mesh recovery system (DS-Seq). By utilizing a higher-precision 2D feature detection framework and leveraging a motion constraint framework based on temporal features to equally incorporate information from past and future frames, our DS-Seq can reconstruct smooth 3D human motion. Additionally, the precision of 3D poses and shapes per frame achieved by our DS-Seq surpasses that of the current state-of-the-art method TCMR by over 10%.