CHMF: Colorful Human Reconstruction Based on Mesh Features

Restoring the color of the 3D human body is a critical issue when reconstructing a 3D human body model from a single-view image. Nonetheless, previous methods have not made use of the reconstructed 3D human body features to aid color restoration, leading to less-than-optimal results. We propose a new method called CHMF in this paper. Unlike previous works, this method fully exploits the inherent features of the 3D human body for color restoration. To achieve this, we first design a deep fix network to extract structural features of the 3D human body to repair color features extracted from single-view image. Then, we utilize a mesh encoding network to extract the shape features of the 3D human body, combining the repaired color features with the transformed shape features to compute the final color. The 3D human body reconstructed by this method not only preserves excellent texture details in terms of color but also eliminates color errors caused by limb occlusion due to human movement, a limitation in existing methods. We evaluate our method on the Thuman2.0 dataset, and extensive experiments have shown that our approach outperforms previous methods both qualitatively and quantitatively.