An Image-Based Method for 3D Human Shapes Retrieval

摘要

Automatically retrieving 3D human shapes from a single 2D image is a challenging problem. The 2D nature of one image makes it difficult to infer depth, pose and style. We propose a novel method for 3D human shape retrieval based on a single image. We present a single-network approach for keypoints detection, which entails simultaneous localization of internal region keypoints and the outer contour keypoints. The network is trained by using multi-task learning, which can handle scale differences between body/foot and face/hand keypoints through an improved architecture. Based on the keypoints, we can estimate the 3D pose, which is used for 3D pose retrieval. From the outer contour keypoints, the 2D closed boundary curve can be automatically generated. We formulate the 2D curve to 3D human shapes similarity calculation as an energy minimization problem for more sophisticated retrieval. Experimental results show that our method can achieve satisfactory retrieval performance on the two benchmark datasets.