Adaptive Surface Fitting Convolution for 3D Point Cloud Analysis

Traditional Convolutional Neural Networks (CNN) are limited to extract informative local features of point clouds due to the fixed geometric structures in convolution kernel against irregular and unstructured point clouds. It usually requires data transformation such as voxelization or projection, inducing a possible loss of information. Instead of fitting the input points to the kernel by regularization, we choose to fit the kernel to input points to conduct convolution. In this paper, we present a new method to define and compute convolution directly on 3D point clouds by Adaptive Surface Fitting Convolution (ASFConv). The key idea is to utilize a set of kernel points distributed on the tangent plane and project them back to point cloud surface. After adapting to the distribution of input points, ASFConv kernel can better capture local neighborhood geometry and benefit the feature extraction. In the experiments, we evaluate our network on two public datasets: ModelNet40 and ShapeNet for classification and segmentation. The experimental results show that our method obtain competitive performances compared to the state-of-the-art.