This research paper presents a novel approach to neural surface reconstruction that incorporates the Truncated Signed Distance Field (TSDF) to substantially reduce the number of samplings required. Traditional techniques are slow due to dense sampling needed to maintain rendering quality. In contrast, this approach leverages the TSDF volume generated only by the trained views, providing a reasonable bound on the sampling from upcoming novel views. The results show an 11-fold increase in inference speed without compromising performance. This method can be robustly plug-and-play into a diverse array of neural surface field models that use the volume rendering technique.
Publication date: 29 Nov 2023
Project Page: https://tsdf-sampling.github.io/
Paper: https://arxiv.org/pdf/2311.17878
