Depth and Frequency Optimized Neural Radiance Field

We introduce DD-NeRF, a sparse-view neural radiance field that utilizes depth constraints for photorealistic rendering with fewer images. By leveraging sparse depth data from the Structure-from-Motion (SFM) stage and generating dense depth via a monocular deep network, we constrain both depth and RGB values during the NeRF rendering process. A dynamically adjusted frequency mask is also used to prioritize low-frequency information early in training, minimizing noise from limited input images. Our method outperforms existing approaches on the LLFF and DTU datasets and enhances other NeRF models with depth constraints for better scene reconstruction.