Latent Partition Implicit with Surface Codes for 3D Representation

Deep implicit functions have shown remarkable shape modeling ability in various 3D computer vision tasks. One drawback is that it hard for them to represent a as multiple parts. Current solutions learn primitives and blend the directly spatial space, which still struggle approximate accurately. To resolve this problem, we introduce novel representation single set of parts latent towards both highly accurate plausibly interpretable modeling. Our insight here part learning blending can be conducted much easier space than space. We name our method Latent Partition Implicit (LPI), because its casting global into local modeling, partitions unity. LPI represents Signed Distance Functions (SDFs) using surface codes. Each code representing whose center on surface, enables us flexibly employ intrinsic attributes shapes or additional properties. Eventually, reconstruct shape, are plausible meshes. multi-level representation, partition different numbers after training. learned without ground truth signed distances, point normals any supervision partition. outperforms latest methods under widely used benchmarks terms reconstruction accuracy interpretability. code, data models available at https://github.com/chenchao15/LPI .