부산대학교 도서관

Implicit neural networks based on point convolution and voxel convolution have been successfully used for point cloud surface reconstruction. This paper proposes an adaptive fusion dual-branch convolutional occupancy network, focusing on the inefficiency of searching local surface based on point convolution technology in reconstruction and the roughness of generated volumetric features based on voxel convolution technology. Dual-branch convolution is used during the encoding stage of the point cloud to extract informative volumetric features. Specifically, the point convolution branch calculates the features of each point in the input point cloud, and the voxel convolution branch calculates the features of each voxel in the regular space of the point cloud. The features of the two branches are then combined using an adaptive fusion module. We performed trilinear interpolation on the feature volume. Experiments on the scene-level Synthetic Rooms dataset and the object-level ShapeNet dataset show that our approach outperforms other methods on most classical metrics.