Parametric Human Shape Reconstruction via Bidirectional Silhouette Guidance

Shuang Sun, Chen Li, Zhenhua Guo, Yuwing Tai

Abstract: We present a method to reconstruct the body geometry of a person by aligning the skinned multi-person linear (SMPL) model to an unconstrained human image. In contrast to previous methods that regress the model parameters from a shared image feature, we decouple the regression of pose and shape parameters in two sub-networks so that we can use different backbone architectures to extract better and more specific features for each regression task while allowing the two sub-networks to work together by our final training loss. We have further proposed a novel bidirectional silhouette constraint to restrict the estimated body geometry. The silhouette constraint is weighted adaptively according to the accuracy of pose estimation in order to handle truncations, occlusions and complex human poses. Experimental results on Human3.6M and UP datasets show that our method outperforms state-of-the-art methods and fits the body segmentation better, especially under extreme human pose conditions.