Fine-grained pillar encoder with spatial adaptation for lidar-based 3D object detection

Abstract

Point clouds based object detection is a fundamental and challenge task in three-dimensional computer vision. However, existing point-based methods struggle to meet real-time object recognition requirements due to excessive raw point counts, whereas pillar-based approaches are constrained by point cloud aggregation algorithms in achieving fine-grained representations. We introduce a novel Spatial Dynamics Adaptation Pillar Network (SDAPNet), which unifies feature saliency enhancement, positional encoding, and adaptive context modeling through a unified computational paradigm, to generate fine-grained pillar from raw point clouds. Specifically, the fine-grained pillar encoding with spatial adaptation incorporates spatial coordinates through focus pooling with position embedding, thereby amplifying salient features in point clouds. Concurrently, its dilation-adaptive 1D convolution dynamically modulates receptive fields to achieve spatial-equivariant representations for 3D object detection. Comprehensive evaluations performed on the KITTI and ONCE datasets conclusively demonstrate the advantages of our method.