Distilling complementary information from temporal context for enhancing human appearance in human-specific NeRF

Abstract Reconstructing and animating digital avatars with free views from monocular videos have been an interesting research task in the computer vision field for a long time. Recently, some methods have introduced a novel category method of leveraging the neural radiance field to represent the human body in a canonical space with the help of the SMPL model. With the deformation of the points from an observation space into a canonical space, the human appearance can be learned in various poses and viewpoints. However, previous methods highly rely on pose-dependent representation learned from frame-independent optimization and ignore the temporal contexts across the continuous motion video, causing a bad influence on the dynamic appearance texture generation. To overcome these problems, we propose a novel free-viewpoint rendering framework, TMIHuman. It aims at introducing temporal information into NeRF-based rendering and distilling task-relevant information from complex pixel-wise representations. To be specific, we build a temporal fusion encoder that imports timestamps into the learning of non-rigid deformation and fuses the visual features of other frames into human representation. Then, we propose to disentangle the fused features and extract useful visual cues via mutual information objectives. We have extensively evaluated our method and achieved state-of-the-art performance on different public datasets.