Dynamic 3D Human Body Shape and Pose Estimation from Multiple Views Based on DQN

Existing multi-view methods generally focus on static body shape and pose estimation. This paper introduces a deep reinforcement learning (DRL) based dynamic estimation framework DQN-MvModeler from multiple views. The DRL method DQN empowers this system to choose the best viewpoints for the optimal poses and body shapes. It also includes a Pose and Shape Decoupled Optimization module (PSDO) to compute the body shapes and poses separately, so that the shapes can be gradually refined with the poses dynamically updated with the spatial-temporal consistency. Experimental results demonstrate the effectiveness of DQN-MvModeler for the robust shape and pose estimation.