Self-supervised depth-guided few-shot neural rendering

Novel views synthesis is an important topic in metaverse application. Existing methods suffer the tremendous training views to guarantee the synthesis quality, which is a stringent condition in practice. To address this problem, we propose a depth-guided and self-supervised method to achieving novel views synthesis in challenging sparse training views. For achieving this goal, we propose a depth information digging strategy and an uncertainty-based depth supervision method. We conduct series of experiments on DTU dataset to demonstrate the rationality of our design. And experiment results represent that our method achieves non-trivial improvement comparing baselines.