Improving 3D Reconstruction from Limited Samples: A Depth-Guided NeRF Approach

In the field of 3D reconstruction, traditional Neural Radiance Fields (NeRF) models face challenges in reconstruction quality when confronted with an insufficient number of samples. This paper introduces an enhanced NeRF model that incorporates a depth estimation network to obtain more continuous depth information, which is then utilized to guide the sampling process. Furthermore, the model incorporates embedding vectors to simulate varying lighting conditions, thereby enhancing its adaptability to changes in illumination. Experiments were conducted on both the LLFF real-world dataset and a custom dataset, and the results demonstrate significant improvements in reconstruction quality under low-sample scenarios, with minor enhancements also observed in scenarios with an adequate number of samples.