2019/03/27 by Siddharth Srivastava, Srivastava, Siddharth, Frédéric Jurie +3
Computer Science · Engineering · #Advanced Neural Network Applications #Autonomous Vehicle Technology and Safety #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Medical Imaging and Analysis
paper · pdf · doi:10.48550/arxiv.1904.08494
openalex publication_date 2019/03/27 · openalex created_date 2022/07/24 · openalex updated_date 2026/07/28
We address the problem of 3D object detection from 2D monocular images in\nautonomous driving scenarios. We propose to lift the 2D images to 3D\nrepresentations using learned neural networks and leverage existing networks\nworking directly on 3D data to perform 3D object detection and localization. We\nshow that, with carefully designed training mechanism and automatically\nselected minimally noisy data, such a method is not only feasible, but gives\nhigher results than many methods working on actual 3D inputs acquired from\nphysical sensors. On the challenging KITTI benchmark, we show that our 2D to 3D\nlifted method outperforms many recent competitive 3D networks while\nsignificantly outperforming previous state-of-the-art for 3D detection from\nmonocular images. We also show that a late fusion of the output of the network\ntrained on generated 3D images, with that trained on real 3D images, improves\nperformance. We find the results very interesting and argue that such a method\ncould serve as a highly reliable backup in case of malfunction of expensive 3D\nsensors, if not potentially making them redundant, at least in the case of low\nhuman injury risk autonomous navigation scenarios like warehouse automation.\n