Sphere-Guided Training of Neural Implicit Surfaces

Figure 1 . We propose a new hybrid approach for learning neural implicit surfaces from multi-view images. In previous methods, the volumetric ray marching training procedure is applied for the whole bounding sphere of the scene (middle left). Instead, we train a coarse sphere-based surface reconstruction (middle right) alongside the neural surface to guide the ray sampling and ray marching. As a result, our method achieves an increased sampling efficiency by pruning empty scene space and better quality of reconstructions (right).