Human Novel View Synthesis from Monocular RGB Video

The novel view synthesis of human bodies is an important step forward for virtual reality. In particular, it is able to predict the invisible domain of moving human bodies, achieving robust and stable reconstruction performance. We propose a novel view synthesis method for human body based on monocular RGB video, while rendering a 360° novel view, avoiding expensive sensor equipment. We integrated the observation space and the canonical space. On one hand, the model could be better optimized and controlled by decomposing geometric human posture and appearance; on the other hand, the rigid and non-rigid motion could be further decoupled to facilitate better optimization approximation. At the same time, we proposed a compact feature description, which could extract the representation to the maximum to achieve better results. We also present qualitative and quantitative results on the performance of the ZJU-MoCap dataset.