PROJECT TITLE :

Detailed Avatar Recovery from Single Image

ABSTRACT:

In this paper, a novel framework for recovering detailed avatar information from a single image is presented. Variations in human shapes, body poses, textures, and viewpoints make it a difficult task to complete. These factors contribute to the difficulty of the task. In the past, techniques typically attempted to reconstruct the shape of a human body by employing a parametric-based template that was deficient in surface details. Because of this, the resulting body shape gives the impression of being naked. In this paper, we propose a novel learning-based framework that combines the resiliency of a parametric model with the adaptability of free-form 3D deformation. Our framework is described in detail in the following paragraphs. Utilizing the constraints provided by body joints, silhouettes, and per-pixel shading information, we train Deep Neural Networks to refine the 3D shape within a Hierarchical Mesh Deformation (HMD) framework. Our approach can restore intricate human body shapes along with their full textures, going beyond the capabilities of skinned models. Experiments have shown that our method is more accurate than other state-of-the-art approaches, both in terms of the number of IoUs in two dimensions and the metric distance between points in three dimensions.