Improved Interactive Reshaping of Humans in Images

Abstract

In this paper, we present an interactive and flexible approach for realistic reshaping of human bodies in a single image. For reshaping, a user specifies a set of semantic attributes like weight and height. Then we use a 3Dmorphable model based image retouching technique for global reshaping of the human bodies in the image such that they satisfy the semantic constraints specified by the user. We address the problem of deformation of the environment surrounding the human body being reshaped, which produces visible artifacts, especially noticeable at regions with structural features, in prior work. We are able to separate the human figure from the background. This allows us to reshape the figure, while preserving the background. Missing regions in the background are inpainted in a manner that maintains structural details. We also provide a quantitative measure for distortion and compare our results with the prior work.