HybridPillars: Hybrid Point-Pillar Network for Real-Time Two-Stage 3-D Object Detection

LiDAR-based 3-D object detection is an important perceptual task in various fields such as intelligent transportation, autonomous driving, and robotics. Existing two-stage point-voxel methods contribute to the boost of accuracy on 3-D object detection by utilizing precise pointwise features to refine 3-D proposals. Although obtaining promising results, these methods are not suitable for real-time applications. First, the inference speed of existing point-voxel hybrid frameworks is slow because the acquisition of point features from voxel features consumes a lot of time. Second, existing point-voxel methods rely on 3-D convolution for voxel feature learning, which increases the difficulty of deployment on embedded computing platforms. To address these issues, we propose a real-time two-stage detection network, named HybridPillars. We first propose a novel hybrid framework by integrating a point feature encoder into a point-pillar pipeline efficiently. By combining point-based and pillar-based networks, our method can discard 3-D convolution to reduce computational complexity. Furthermore, we propose a novel pillar feature aggregation network to efficiently extract bird's eye view (BEV) features from pointwise features, thereby significantly enhancing the performance of our network. Extensive experiments demonstrate that our proposed HybridPillars not only boosts the inference speed, but also achieves competitive detection performance compared to other methods. The code will be available at https://github.com/huangzhicong3/HybridPillars.