Seam-Aware Rendering Quality Enhancement Network For Compressed 3D Scene.

Typically, UV maps generated from real-world reconstructed meshes tend to have extensive fragmentation, i.e., numerous UV charts. This fragmentation of the UV map results in severe discontinuities in the texture map, which degrades texture map compression performance and causes texture bleeding artifacts near the seams of the rendered images of the compressed texture map. To solve this problem, this paper introduces a convolutional neural network (CNN)-based post-processing method for enhancing the rendering result of compressed texture map, especially focusing on reducing bleed artifacts. The CNN architecture of the proposed method is designed to perform both rendering quality enhancement and texture bleed artifact removal by being aware of seams in the rendered image. Through the proposed post-processing method, rendered images have a lower distortion in terms of rendering quality, consequently, this enhancement leads to better compression performance in terms of bitrate and rendering distortion. Our experiments show average bitrate savings of 5.69% and 14.1% in PSNR and MS-SSIM with proposed method compared to where it is not performed, and it also notably reduce texture bleed artifacts.