Adaptive octree 3D image reconstruction based on plane patch

In this study, an adaptive octree convolutional neural network based on plane patches is proposed for effective 3D shape encoding and decoding. Unlike volume-based or octree-based convolutional neural network (CNN) methods, which represent 3D shapes with the same voxel resolution, the proposed method can use planes and adaptively represent the 3D shapes of octree nodes with different levels. The patch models the 3D shape within each octree node, whereby the patch-based adaptive representation is utilized in the proposed adaptive patch octree convolutional neural network (O-CNN) encoder and decoder for the encoding and decoding of 3D shapes. The adaptive patch O-CNN encoder takes the plane patch normal and displacement as input and performs three-dimensional convolution on the octree nodes of each level, whereas the adaptive patch O-CNN decoder infers each level. The shape occupancy rate and subdivision state of the octree node as well as the best plane normal and displacement of each leaf octree node are estimated. As a general framework for 3D shape analysis and generation, adaptive patch O-CNN not only reduces memory and computational costs but also exhibits better shape generation capabilities than existing 3D-CNN methods. Shape prediction is performed using a single image to verify the efficiency and effectiveness of the generation task of the adaptive O-CNN. The chamfer distance error is 0.274, which is lower than that of OctGen (0.294), resulting in a better reconstruction effect. © 2022, Science Press. All right reserved.