2018/11/09 by Kiwoo Shin, Shin, Kiwoo, Youngwook Paul Kwon +3 · 2 citations
Computer Science · Engineering · #Advanced Neural Network Applications #Autonomous Vehicle Technology and Safety #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Industrial Vision Systems and Defect Detection
paper · pdf · doi:10.48550/arxiv.1811.03818
openalex publication_date 2018/11/09 · openalex created_date 2019/11/22 · openalex updated_date 2026/07/28
We present RoarNet, a new approach for 3D object detection from a 2D image\nand 3D Lidar point clouds. Based on two-stage object detection framework with\nPointNet as our backbone network, we suggest several novel ideas to improve 3D\nobject detection performance. The first part of our method, RoarNet2D,\nestimates the 3D poses of objects from a monocular image, which approximates\nwhere to examine further, and derives multiple candidates that are\ngeometrically feasible. This step significantly narrows down feasible 3D\nregions, which otherwise requires demanding processing of 3D point clouds in a\nhuge search space. Then the second part, RoarNet3D, takes the candidate\nregions and conducts in-depth inferences to conclude final poses in a recursive\nmanner. Inspired by PointNet, RoarNet3D processes 3D point clouds directly\nwithout any loss of data, leading to precise detection. We evaluate our method\nin KITTI, a 3D object detection benchmark. Our result shows that RoarNet has\nsuperior performance to state-of-the-art methods that are publicly available.\nRemarkably, RoarNet also outperforms state-of-the-art methods even in settings\nwhere Lidar and camera are not time synchronized, which is practically\nimportant for actual driving environments. RoarNet is implemented in Tensorflow\nand publicly available with pre-trained models.\n