PSIDet: Probabilistic Structure Information from Point Cloud for 3D Object Detection

Abstract

Detection of 3D scenes from LIDAR point cloud is a challenging task. Recent studies have shown that the performance of 3D detectors degrades dramatically due to unbalanced point density and missing points. In this paper, we present a P robabilistic S tructure I nformation Det ection network, a general approach to enhance the structure information for feature representation. Our key focus is on extracting the structure feature and combining it with original feature. Specifically, We propose a plug- and-play module, where boundary information can be nearly cost-free extracted because the feature is shared with that encoded by the backbone network. Also, to maximize the use of extracted structural information, we design a Weighted Boundary Prediction (WBP) Module, aiming to encourage the detector pay more attention to the structure information of the object. By merging the structure feature with the original feature, we obtain an augmented feature representation, which can be directly used by the second stage of the detector. Extensive experiments show our PSIDet significantly improves the result on Car 3D detection for the KITTI benchmark.

Keywords: 3D object detection, Point cloud, Structure information, LiDAR, Convolutional neural network, Autonomous driving.