
摘要
我们提出了一种新颖且高性能的3D目标检测框架,命名为PointVoxel-RCNN(PV-RCNN),用于从点云中进行精确的3D目标检测。所提出的方法深入结合了3D体素卷积神经网络(CNN)和基于PointNet的集合抽象,以学习更具区分性的点云特征。该方法充分利用了3D体素CNN高效的学习能力和高质量的候选框生成能力,以及基于PointNet的网络灵活的感受野。具体而言,所提出的框架通过一种新型的体素集合抽象模块,利用3D体素CNN将3D场景总结为少量的关键点,从而节省后续计算并编码代表性场景特征。鉴于由体素CNN生成的高质量3D候选框,我们提出了RoI网格池化方法,通过关键点集合抽象和多个感受野将提案特定特征从关键点抽象到RoI网格点。与传统的池化操作相比,RoI网格特征点编码了更丰富的上下文信息,有助于准确估计目标置信度和位置。在KITTI数据集和Waymo开放数据集上的大量实验表明,仅使用点云的情况下,我们的PV-RCNN显著超越了现有的最先进3D检测方法。代码可在https://github.com/open-mmlab/OpenPCDet 获取。
代码仓库
sshaoshuai/PV-RCNN
pytorch
GitHub 中提及
jhultman/PV-RCNN
pytorch
GitHub 中提及
KangchengLiu/FAC_Foreground_Aware_Contrast
pytorch
GitHub 中提及
KangchengLiu/RM3D
pytorch
GitHub 中提及
KPeng9510/MASS
pytorch
GitHub 中提及
sunshenggu/xc_eval_pcdet
pytorch
GitHub 中提及
open-mmlab/OpenPCDet
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| 3d-object-detection-on-kitti-cars-easy | PV-RCNN | AP: 90.25% |
| 3d-object-detection-on-kitti-cars-hard | PV-RCNN | AP: 76.82% |
| 3d-object-detection-on-kitti-cyclists | PV-RCNN | AP: 63.71% |
| 3d-object-detection-on-kitti-cyclists-easy | PV-RCNN | AP: 78.60% |
| 3d-object-detection-on-kitti-cyclists-hard | PV-RCNN | AP: 57.65% |
| 3d-object-detection-on-waymo-all-ns | PV-RCNN | APH/L2: 71.52 |
| 3d-object-detection-on-waymo-cyclist | PV-RCNN | APH/L2: 71.16 |
| 3d-object-detection-on-waymo-pedestrian | PV-RCNN | APH/L2: 70.16 |
| 3d-object-detection-on-waymo-vehicle | PV-RCNN | APH/L2: 73.23 |
| birds-eye-view-object-detection-on-kitti | PV-RCNN | AP: 68.89% |
| birds-eye-view-object-detection-on-kitti-8 | PV-RCNN | AP: 82.49 |
| birds-eye-view-object-detection-on-kitti-9 | PV-RCNN | AP: 62.41 |
| birds-eye-view-object-detection-on-kitti-cars | PV-RCNN | AP: 90.65% |
| birds-eye-view-object-detection-on-kitti-cars-4 | PV-RCNN | AP: 94.98 |
| birds-eye-view-object-detection-on-kitti-cars-5 | PV-RCNN | AP: 86.14 |
| robust-3d-object-detection-on-kitti-c | PV-RCNN | mean Corruption Error (mCE): 90.04% |