CenterNet3D: An Anchor Free Object Detector for Point Cloud

Accurate and fast 3D object detection from point clouds is a key task in autonomous driving. Existing one-stage methods can achieve real-time performance, however, they are dominated by anchor-based detectors which inefficient require additional post-processing. In this paper, we eliminate anchors model an as single point—the center of its bounding box. Based on the point, propose anchor-free CenterNet3D network that performs without anchors. Our uses keypoint estimation to find points directly regresses boxes. However, because inherent sparsity clouds, likely be empty space makes it difficult estimate accurate boundaries. To solve issue, extra corner attention module enforce CNN backbone pay more Besides, considering suffer discordance between predicted boxes corresponding classification confidences, develop efficient keypoint-sensitive warping operation align confidences proposed non-maximum suppression free simpler. We evaluate widely used KITTI dataset challenging nuScenes dataset. method outperforms all state-of-the-art has comparable performance two-stage well. It inference speed 20 FPS achieves best accuracy trade-off. source code will released at https://github.com/wangguojun2018/CenterNet3d .