Monocular Human Body Shape Estimation: A Generation-aid Approach

Observing human beings from monocular images is one of the basic tasks of computer vision. Reconstructing human bodies from monocular images mainly includes the reconstruction of posture and body shape. However, in the past studies, researchers were more interested in pose estimation, ignoring the study of body shape, and this paper focuses on the estimation of the body shape of a 3D model. Learning body parameters via instance segmentation requires a large number of labels. While the parameters based on pose estimation are completely based on the results of key points detection, which effect is not friendly for pictures with poor angles and low resolution. In response to the above problems, we propose a method to automatically generate datasets. The dataset provides low-resolution images and labels of various angles and blurred shapes. On the generated low-resolution and poorly angled dataset, we propose a generative-assisted deep learning network framework. Experiments show that the framework can effectively estimate the body shape parameters of the model from monocular images.