3D Object Detection Based on Polar Representation for Better Comprehensive Performances

Abstract

Multi-modal 3D object detection is an important task in autonomous driving systems, where cameras and LiDAR provide complementary semantic and geometric information. Most existing BEV fusion methods are designed based on the Cartesian representation space, which does not fully match the sensing geometry of camera and LiDAR. This generally leads to redundant computation in distant regions. To address this issue, GARF, a geometry-aware polar BEV framework, is presented for multi-modal 3D object detection. GARF organizes camera and LiDAR features in a unified polar BEV space, which can represent spatial resolution more compactly. For the camera branch, the uncertainty-guided transformation of the polar view is designed to improve the reliability of depth estimation. Then, the generated polar BEV feature is further refined to attenuate radial noise and angular discontinuity. For the LiDAR branch, the polar-aware sparse feature extraction and distortion correction modules are designed to deal with the anisotropic structure and geometric distortion caused by polar voxelization. For multi-modal fusion, the region-aware cross-modal fusion strategy and polar detection head with anisotropic Gaussian center response map are developed, which achieve effective feature interaction and consistent geometry supervision. The experimental results on nuScenes show that GARF achieves 71.8% mAP and 73.7% NDS, improving the baseline by 3.3% mAP and 2.3% NDS. Meanwhile, the inference speed increases from 7.1 FPS to 8.9 FPS, and the consumption of GPU memory decreases from 41,114 MiB to 33,346 MiB.