Learning Pose Grammar to Encode Human Body Configuration for 3D Pose Estimation

In this paper, we propose a knowledge-guided pose grammar network to tackle the problem of 3D human pose estimation. Our model directly takes 2D poses as inputs and learns the generalized 2D-3D mapping function, which renders high applicability. The proposed network consists of a base network which efficiently captures pose-aligned features and a hierarchy of Bidirectional RNNs on top of it to explicitly incorporate a set of knowledge (e.g., kinematics, symmetry, motor coordination) and thus enforce high-level constraints over human poses. In learning, we develop a pose-guided sample simulator to augment training samples in virtual camera views, which further improves the generalization ability of our model. We validate our method on public 3D human pose benchmarks and propose a new evaluation protocol working on cross-view setting to verify the generalization ability of different methods. We empirically observe that most state-of-the-arts face difficulty under such setting while our method obtains superior performance.