AE-IA-SSD: an attention-enhanced point-based detectors for 3D LiDAR point clouds

We study the efficient method of extracting crucial information from LiDAR point clouds for 3D Object Detection. Some prior works have demonstrated the importance of downsampling algorithms for 3D Object Detection, that is, the results of downsampling determine the performance of the algorithm to a large extent. Inspired by this, we introduce AE-IA-SSD, an efficient, single-stage, point-based 3D detector. We propose an importance-aware downsampling strategy by incorporating an attention mechanism at the subsampling layer to refine the selection process of keypoints. Specifically, a learnable attention module is used to dynamically weight the importance of each point at the secondary sampling stage, guided by local and global context, which enables the allocation of computational resources to point cloud regions that are more relevant to the object detection task, resulting in a significant improvement in detection accuracy while maintaining or even reducing computational footprint. Additionally, to improve the efficiency of the algorithm, We designed AE-IASSD using an encoder-only architecture to streamline its structure for efficiency. We evaluated our method on the widely used KITTI dataset, demonstrating substantial improvements over prior models.