AtlantaSDF: Neural 3D Indoor Scene Reconstruction with the Atlanta-world Assumption

This paper primarily focuses on resolving the issue of reconstructing weakly textured areas in indoor environments in multi-view 3D reconstruction. Previous works could yield impressive reconstruction results on objects with textured areas but have encountered difficulties in low-texture planar areas commonly found indoors. In this paper, plane constraints are applied to 3D reconstruction methods based on implicit neural representation. We use an MLP network to represent scene geometry by using signed distance functions(SDF). For the floor and wall regions that have been predicted by the 2D semantic segmentation network, the geometric normal vectors of the above regions are regularized using planar constraints based on the Atlanta-world assumption. The result of semantic segmentation affects the accuracy of 3D reconstruction, the semantic logits of a point are calculated in 3D space and rendered into 2D space to get the probability that the point belongs to each region to optimize the semantic segmentation, and at the same time, weighting it to the loss function also optimizes the geometric result of 3D reconstruction, based on which a collaborative optimization strategy is proposed. Experimental results have shown that this method significantly outperforms previous methods in the quality of 3D reconstruction in weakly textured areas.