PointRCNN++: Towards More Accurate Two-Stage 3D Object Detection from Point Cloud

Abstract

With significant potential in autonomous driving, and robotics, 3D object detection has garnered increasing attention among researchers. Leveraging its reliability in depth information, lidar stands out as one of the pivotal sensors in the realm of autonomous driving. PointRCNN, a point-based twostage object detector tailored for point cloud, has demonstrated commendable performance on many benchmarks though exhibiting certain limitations. Before being fed into neural network, raw point cloud undergoes down-sampling procedure, which potentially results in the loss of foreground points. Recognizing the non-uniform nature of point cloud distribution, we introduce an enhanced sampling approach termed distance bin-based FPS. This novel method adaptively adjusts the sampling rate based on the proximity of points, allocating lower rates for dense nearby points and higher rates for sparser distant points, in order to retain more foreground points within small or distant objects. Furthermore, we incorporate Point-Attention into the PointNet++ backbone to capture long-range relationships within point clouds. By combining the distance bin-based FPS and an enhanced backbone, our proposed model PointRCNN++ exhibits improved perceptual capabilities, facilitating more precise object detection. Experimental results on the KITTI dataset substantiate our model's effectiveness, showcasing a substantial performance enhancement over the baseline, thus validating the efficacy of our proposed enhancements.