SAMUNet: Enhancing pillar-based 3D object detection in autonomous driving with Shape-aware Mini-Unet

Pillar-based 3D object detection methods outperform traditional point-based and voxel-based methods in terms of speed. However, existing methods struggle with accurately detecting large objects in complex environments due to the limitations in capturing global spatial dependencies. To address these issues, this paper proposes Shape-aware Mini-Unet Network (SAMUNet), a simple yet effective hierarchical 3D object detection network. SAMUNet incorporates multiple Sparse Mini-Unet blocks and a Shape-aware Center Head. Concretely, after converting the original point cloud into pillars, we first progressively reduce the spatial distance between distant features through downsampling in the Sparse Mini-Unet block. Then, we recover lost details through multi-scale feature fusion, enhancing the model’s ability to detect various objects. Unlike other methods, the upsampling operation in the Sparse Mini-Unet block only processes the effective feature coverage area of the intermediate feature map, significantly reducing computational costs. Finally, to further improve the accuracy of bounding box regression, we introduce Shape-aware Center Head, which models the geometric information of the bounding box’s offset direction and scale using 3D Shape-aware IoU. Extensive experiments on the nuScenes and Waymo datasets demonstrate that SAMUNet excels in detecting large objects and overall outperforms current state-of-the-art detectors, achieving 72.0% NDS and 67.7% mAP.