FSS-Net: A Fast Search Structure for 3D Point Clouds in Deep Learning
Jiawei Wang, Yan Zhuang, Yisha Liu
Abstract
The deep learning methods achieve good results in the semantic segmentation and classification of the 3D point clouds. The popular convolutional neural networks illustrate the importance of using the neighboring information of the points. Searching the neighboring points is an important process that can get the context information of each point. The K-nearest neighbor (KNN) search and ball query methods are usually used to find the neighboring points, but a long time is required to construct the KD-tree and calculate the Euclidean distance. In this work, we introduce a fast approach (called the voxel search) to finding the neighbors, where the key is to use the voxel coordinates to search the neighbors directly. However, it is difficult to apply this method directly to the general network structure such as the U-net. In order to improve its applicability, the corresponding up-sampling and down-sampling methods are proposed. Additionally, we propose a fast search structure network (FSS-net) which consists of the feature extraction layer and the sampling layer. In order to demonstrate the effectiveness of the FSS-net, we conduct experiments on a single object in both indoor and outdoor environments. The speed of the voxel search approach is compared with that of the KNN and ball query. The experimental results show that our method is faster and can be directly applied to any point-based deep learning networks.