Adaptive division of multi-layer feature fusion for detailed 3D reconstruction

Point cloud is a versatile and easily accessible lightweight data representation, is widely used in visual processing. However, the discrete nature of 3D objects, along with the challenges posed by noise interference and uneven density, contribute to the complexity of accurately reconstructing 3D objects. Due to only using global features for 3D reconstruction makes the result surface difficult to describe the target details, and only using local features for 3D reconstruction makes the result surface discontinuous. In this paper, we propose a method of adaptive division and multilayer feature fusion to reconstruction the three-dimensional shapes with unoriented point clouds. We adjusted the octree division rule based on neural network to adaptively divide the target point cloud with a more reasonable division method. In addition, we inherit the moving least squares method based on the octree and transform the discrete point set into a three-dimensional surface. The difference is that we combine the octree features of different depths to obtain the context information of the shape in a local-plus-global manner. We conduct comparative experiments with other learning-based methods on ShapeNet dataset, our method outperforms advanced learning-based surface reconstruction approaches, shows superior quantitative and qualitative results.