Frustum consistency augmentation for 3D object detection in LiDAR

Abstract In autonomous driving, LiDAR is essential for its high precision and reliability. The identification of objects that are either far away or partially obscured continues to pose significant difficulties, primarily because of the limited density and resolution of point cloud information. While combining image data or using virtual points from depth completion helps, it adds significant computational overhead. In this paper, we innovatively implement a combination of 2D detectors and raw points within the RoI (region of interest) to filter virtual points to resolve the challenges previously outlined. When the model identifies where to focus, we hope it can discern object contours and subsequently predict more accurate bounding boxes. To achieve this, we introduce the Contour-Aware Augmentation module which enhances target contours by mirror-replicating point clouds within the ground-truth bounding boxes during training. After that, it strategically discards point clouds to effectively simulate various distances and occlusion scenarios. Additionally, we implement a consistency loss model to align predictions from both raw and virtual point clouds, fostering synergistic interactions among the modules. These advancements significantly enhance the effectiveness of our method on the KITTI benchmark, resulting in a 3.43% improvement in 3D detection accuracy in the hard category compared to the baseline.