CG-SSD: Corner guided single stage 3D object detection from LiDAR point cloud

Detecting accurate 3D bounding boxes of the object from point clouds is a major task in autonomous driving perception. At present, anchor-based or anchor-free models that use LiDAR for detection center assigner strategy to infer boxes. However, real-world scene, due occlusions and effective range system, only part surface can be covered by collected clouds, there are no measured points corresponding physical center. Obtaining aggregating incomplete will bring loss accuracy direction dimension estimation. To address this problem, we propose corner-guided single-stage model (CG-SSD). Firstly, within single frame assigned regular grids. sparse convolution backbone network composed residual layers sub-manifold convolutional used construct bird’s eye view (BEV) features further deeper feature mining lite U-shaped network; Secondly, novel auxiliary module (CGAM) with adaptive corner classification algorithm proposed incorporate supervision signals into neural network. CGAM explicitly designed trained estimate locations partially visible invisible corners obtain more representation, especially small partial occluded objects; Finally, deep both networks concatenated fed head predict objects scene. The experiments demonstrate CG-SSD achieves state-of-art performance on ONCE benchmark supervised using cloud data, 62.77% mAP. Additionally, Waymo Open Dataset show extended most which BEV detect objects, as plug-in +1.17%∼+14.23% AP improvement. code available at https://github.com/mrqrs/CG-SSD.