Spatial-Aware Multi-Dimensional Range-Enhanced Networks for Efficient 3D Object Detection

Virtual point-based 3D object detection is a visual detection method that fuses RGB images and LiDAR data through depth completion techniques. However， the high-density point clouds generated by this method can contain up to 600000 detection points per frame， imposing strict efficiency requirements on algorithms during real-time deployment. Existing methods rely on single-dimensional spatial assumptions， treating nearby regions as uniformly redundant and adopting random discarding strategies. Such approaches fail to capture the true redundancy distribution jointly determined by spatial location， semantic category， and feature quality， resulting in inefficient use of computational resources. To address this issue， this study proposes spatial-aware multi-dimensional range-enhanced networks （SMR-3D）. This network constructs a multi-dimensional joint modeling framework that unifies the optimization of spatial， semantic， and range-dependent feature dimensions to improve computational efficiency. At the data level， SMR-3D improves sampling efficiency through spatial-aware sampling （SAS） and multi-dimensional importance scoring sampling （MISS）. SAS partitions the 3D space into near-， mid-， and far-range regions， assigning different voxel discard rates according to distance， with higher discard rates for near-range regions to reduce redundancy and lower discard rates for far-range regions to preserve inherently sparse critical voxels. MISS model semantic parameters of four attributes—reflectance， height， orientation， and distance—and learns the importance of each attribute to prioritize the retention of critical voxels with target-characteristic features. At the feature level， the range-aware network （RangeNet） employs learnable distance bins and a soft assignment mechanism to compensate for feature degradation caused by increasing distance through channel-wise adaptive weighting. Finally， experiments on the KITTI validation set demonstrate that SMR-3D achieves 84.07% 3D average precision （AP） at the moderate difficulty level， representing a 1.72 percentage point improvement over the random sampling baseline， with an inference time of 54 ms. Qualitative analysis shows that SMR-3D successfully detects target vehicles at distant ranges that are missed by the baseline method， achieving a favorable balance between detection accuracy and real-time performance in autonomous driving applications.