<p>LiDAR-based three-dimensional (3D) object detection is crucial in autonomous driving. The existing approach of solely relying on points to compensate for the missing information of voxels is inefficient and inadequate. Although the information contained in reflectance intensity affords significant potential for feature mining, effective network structures and architectural designs for harnessing reflectance information efficiently and ensuring overall efficiency are lacking. To efficiently and comprehensively exploit the information contained in point clouds, we propose a novel paradigm for two-stage 3D object detection based on the “bird’s-eye view (BEV) of reflectivity” and “voxel.” We design an efficient two-dimensional network to extract features from the reflectance intensity BEV, compensating for the information loss caused by voxel downsampling. Moreover, a hierarchical encoding method for reflectance intensity BEV is designed, and features are extracted through the proposed feature pyramid backbone network to focus on learning more discriminative features carried in reflectance intensity, which helps to distinguish objects with similar shapes. Finally, we incorporate the voxel features extracted from the 3D sparse convolutional backbone network with the reflectance intensity BEV features to guide the generation of pre-selected bounding boxes, and the multi-scale voxel pooling features of the regions of interest are fused to further optimize the detection boxes. Experiments performed on the widely used datasets demonstrate the higher detection accuracy of the proposed method than that of mainstream methods.</p>