Neural Mesh Fusion: Unsupervised 3D Planar Surface Understanding

This paper presents Neural Mesh Fusion (NMF), an efficient approach for jointoptimization of polygon mesh from multi-view image observations andunsupervised 3D planar-surface parsing of the scene. In contrast to implicitneural representations, NMF directly learns to deform surface triangle mesh andgenerate an embedding for unsupervised 3D planar segmentation throughgradient-based optimization directly on the surface mesh. The conductedexperiments show that NMF obtains competitive results compared tostate-of-the-art multi-view planar reconstruction, while not requiring anyground-truth 3D or planar supervision. Moreover, NMF is significantly morecomputationally efficient compared to implicit neural rendering-based scenereconstruction approaches.