Depth-Supervised Neural Radiance Fields for Accurate Novel View Synthesis

We present Depth-NeRF as a promising solution to address the limitations of Neural Radiance Fields (NeRF). NeRF effectively encodes scenes into neural representations, allowing for the photo-realistic rendering of novel views. However, it falls short in capturing surface geometry, restricting its rendering capabilities to surface colors alone. For overcoming these limitations, we have undertaken substantial modifications to the NeRF network architecture and strategically incorporated depth information to guide its optimization. This method ensures that the ray termination distribution conforms to the scene's geometric properties, thereby producing renderings with improved precision and accuracy. Compared to NeRF, our method achieves a substantial improvement in rendering quality, while simultaneously reducing the RMSE between the rendered and actual scene depth. These improvements unequivocally demonstrate our approach's superior ability to capture the scene's geometric information.