Efficient Human Rendering with Geometric and Semantic Priors

Abstract

Recently, human rendering has attracted many attention thanks to its vast applications. With new advances in neural rendering and radiance field, synthesizing realistic novel view images from multi-view camera images can be achived with less manual labour. However, due to the data-driven nature of such algorithms, the efficiency for both time and computation can be unsatisfying. Hence, we propose an efficient human rendering pipeline, generating geometric and semantic guidances as priors to further enhance both efficiency and quality. Specifically, a semantic human part parsing guides the pixel sampling in 2D space, and a mesh prior is utilized to guide an occupancy field for effective ray sampling in 3D space. As a result, we achieved considerable improvement over previous methods in both efficiency and rendering quality.