Neural RGB-D Surface Reconstruction
In this work, we explore how to leverage the success of implicit novel view synthesis methods for surface reconstruction. Methods which learn a neural radiance field have shown amazing image synthesis results, but the underlying geometry representation is only a coarse approximation of the real geometry. We demonstrate how depth measurements can be incorporated into the radiance field formulation to produce more detailed and complete reconstruction results than using methods based on either color or depth data alone. In contrast to a density field as the underlying geometry representation, we propose to learn a deep neural network which stores a truncated signed distance field. Using this representation, we show that one can still leverage differentiable volume rendering to estimate color values of the observed images during training to compute a reconstruction loss. This is beneficial for learning the signed distance field in regions with missing depth measurements. Furthermore, we correct for misalignment errors of the camera, improving the overall reconstruction quality. In several experiments, we show-cast our method and compare to existing works on classical RGB-D fusion and learned representations.
We test our method on the ScanNet dataset which provides RGB-D sequences of room-scale scenes. We compare our method to the original ScanNet BundleFusion reconstructions which often suffer from severe camera pose misalignment. Our approach jointly optimizes for the scene representation network as well as the camera poses, leading to substantially reduced misalignment artifacts in the reconstructed geometry.
