Projection-based point convolution for fast point cloud segmentation

Abstract

Understanding point cloud has recently gained huge interests following the development of 3D scanning devices and the accumulation of large-scale 3D data. In this research, we study various point cloud processing algorithms and propose a novel approach that provides possibility to overcome the limitations of existing methods. Specifically, we focus on the speed of point cloud processing algorithms and show that effective algorithms run too slowly to perform real-time analysis, while fast algorithms relatively lack in accuracy on the target task. Based on the observations that point-based and voxel-based 3D convolutional methods show better performance while projection-based methods run faster, we present a novel approach, which is a hybrid method of projection- and point-based algorithms, that leverages the advantage of each method. First, we propose a convolutional module, named Projection-based Point Convolution (PPConv), that uses 2D convolutions and multi-layer perceptrons (MLPs) as its components. In PPConv, point features are processed through two branches: point branch, which consists of MLPs, and projection branch that transforms point features into a 2D feature map and then use 2D convolutions to transform these features. We do not use the time-consuming operations which are used in point-based or voxel-based convolutions while constructing PPConv

thus it has advantage in fast point cloud processing. When combined with a learnable projection and effective feature fusion strategy, PPConv achieves superior efficiency compared to popular point-based and voxel-based methods, even with a simple backbone architecture based on PointNet++. Next, through deeper study on the architecture design, we construct an efficient point cloud processing framework that can operate on larger-scale point clouds. Throughout the experiments, we demonstrate the efficiency of projection-based point convolutional approach in terms of the trade-off between inference time and segmentation performance. We analyze the efficiency of our approach using point clouds of various scales, including object shapes, indoor scenes, and outdoor LiDAR scenes. Through this research, we intend to provide an efficient approach to point cloud processing based on projection, which can further be combined with various research on improving the performance of 2D convolutional architectures or the running speed of 2D CNNs.