Adaptive O-CNN: A Patch-based Deep Representation of 3D Shapes

摘要

We present an Adaptive Octree-based Convolutional Neural Network (AdaptiveO-CNN) for efficient 3D shape encoding and decoding. Different fromvolumetric-based or octree-based CNN methods that represent a 3D shapewith voxels in the same resolution, our method represents a 3D shape adaptivelywith octants at different levels and models the 3D shape within eachoctant with a planar patch. Based on this adaptive patch-based representation,we propose an Adaptive O-CNN encoder and decoder for encoding anddecoding 3D shapes. The Adaptive O-CNN encoder takes the planar patchnormal and displacement as input and performs 3D convolutions only at theoctants at each level, while the Adaptive O-CNN decoder infers the shapeoccupancy and subdivision status of octants at each level and estimates thebest plane normal and displacement for each leaf octant. As a general frameworkfor 3D shape analysis and generation, the Adaptive O-CNN not onlyreduces the memory and computational cost, but also offers better shapegeneration capability than the existing 3D-CNN approaches. We validateAdaptive O-CNN in terms of efficiency and effectiveness on different shapeanalysis and generation tasks, including shape classification, 3D autoencoding,shape prediction from a single image, and shape completion for noisyand incomplete point clouds.