BADet: Boundary-Aware 3D Object Detection from Point Clouds
作者：
Highlights：
• We propose BADet framework which effectively models local boundary correlations of an object in the form of local neighborhood graph, which explicitly facilitates a complete boundary for each individual proposal by the means of an information compensation mechanism.
• We propose a lightweight region feature aggregation module to make use of informative semantic features, leading to significant improvement with manageable memory overheads.
• Our BADet outperforms all previous state-of-the-art methods with remarkable margins on KITTI BEV detection leaderboard and ranks 1st on Car category of Moderate difficulty as of Apr. 17th, 2021. Furthermore, comprehensive experiments are conducted on KITTI Dataset in diverse evaluation settings to analyze the effectiveness of BADet.
摘要
•We propose BADet framework which effectively models local boundary correlations of an object in the form of local neighborhood graph, which explicitly facilitates a complete boundary for each individual proposal by the means of an information compensation mechanism.•We propose a lightweight region feature aggregation module to make use of informative semantic features, leading to significant improvement with manageable memory overheads.•Our BADet outperforms all previous state-of-the-art methods with remarkable margins on KITTI BEV detection leaderboard and ranks 1st on Car category of Moderate difficulty as of Apr. 17th, 2021. Furthermore, comprehensive experiments are conducted on KITTI Dataset in diverse evaluation settings to analyze the effectiveness of BADet.
论文关键词：3D object detection,autonomous driving,graph neural network,boundary aware,point clouds
论文评审过程：Received 2 August 2021, Revised 1 December 2021, Accepted 6 January 2022, Available online 10 January 2022, Version of Record 15 January 2022.
论文官网地址：https://doi.org/10.1016/j.patcog.2022.108524
