Fast Sampling for Neural Surface Field via Truncated Signed Distance Field

Abstract

Neural surface reconstruction from multiple views has demonstrated remarkable performance, but it is hindered by slow inference times compared to traditional methods. This obstacle is largely due to the need for dense sampling to achieve high performance. Previous studies on sampling relied on initial samplings across the entire space. This approach led to performance degradation when the sample count was reduced and necessitated the training of an additional neural network. This thesis presents an innovative method that greatly reduces the number of required samples by incorporating the Truncated Signed Distance Field (TSDF), which is designed to be compatible without modification of the model architecture and can be seamlessly integrated with various neural surface field models. A single-resolution TSDF volume, constructed from the depth maps of training views, excludes empty space and the interior of objects that do not affect rendering as the ray traverses the TSDF voxels. To handle bounds of different lengths, an adaptive sampling method is introduced to adjust the number of samples based on the length of the sample bound, allowing the model to determine the appropriate samples per ray. The proposed method was tested on a large indoor dataset and a synthetic dataset, rendering 11× faster than existing sample methods with no performance loss. Further- more, the method remains effective when applied to a model with relatively poor geometry, where influence points are not focused on the surface.