DKD-NeRF: Depth Knowledge-Distillation NeRF for Sparse Input Views

We propose DKD-NeRF, a novel view synthesis model tailored for sparse-view scenarios, which alternately trains a depth knowledge distillation framework and Neural Radiance Field (NeRF) during training, enabling high-quality novel view generation in indoor sparse-view scenes. In sparse-view scenarios, NeRF struggles to fit correct geometric shapes, leading to a significant decline in the quality of generated novel views. Existing methods leverage depth priors to constrain the NeRF training process, but the coarse depth without scene priors suffers from ambiguity. Our approach alternately updates the depth estimation model and the NeRF, progressively refining the depth values during training. Furthermore, to mitigate the impact of the sparse training set, we propose a strategy called Random Reflected Lights. Experimental results on public datasets and in-the-wild dataset demonstrate that, compared to existing methods, DKD-NeRF can generate higher-quality novel views from a set of indoor sparse views.