LiDAR point clouds is essential but challenging for 3D object detection. Most existing methods focus on classification and bounding box regression in the way of designing effective network to learn expressive feature from this irregular data. However, there exists another bottleneck, which is to have a good detection confidence score to correctly pick out the best matching one from several bounding boxes corresponding to the same ground truth bounding box. In this paper, we propose a novel 3D object detection framework 3D IoU-Net aimed to perceive an accurate detection confidence score, that is, the 3D Intersection-over-Union (IoU) between the predicted bounding box and the corresponding ground truth bounding box. Specifically, the proposed 3D IoU-Net generates proposals and point-wise features from raw point cloud. Give a 3D proposal, the IoU sensitive feature is pooled by our Attentive Corner Aggregation (ACA) module and Corner Geometry Encoding (CGE) module for extra 3D IoU prediction. By considering the visible parts varies from the point cloud gathering angle, the ACA module aggregate the point-wise feature from each corners perspective with different attention to generate a more unified feature. Besides, the novel CGE module encodes the geometric information of bounding box itself, which is always neglected in other methods. In addition, the IoU alignment operation further boosts the 3D IoU prediction. Thanks to accurate 3D IoU prediction value as detection confidence, the better localized bounding boxes are reasonably prevented from being suppressed. Experiments on KITTI car detection benchmark show that 3D IoU-Net achieves state-of-the-art performance. Copyright © 2020, The Authors. All rights reserved.
- Computer Vision and Pattern Recognition (cs.CV)