FasterV-RCNN: Efficient Point Cloud 3D Object Detection Framework

Recent advances in 3D object detection rely heavily on the representation of 3D data. Several high-performance 3D detectors rely on a point-based structure as it preserves precise point positions. However, point-level features incur high computation overheads due to unordered storage. Conversely, the voxel-based structure is better suited for feature extraction but often results in slow inference times as the interaction between points and voxels can be time-consuming. In this work, we propose a new point cloud detection framework that dynamically and adaptively processes raw input point cloud data to achieve higher inference speeds. We conduct extensive experiments on the widely used KITTI dataset, and our results demonstrate that our proposed FasterV-RCNN method achieves higher detection accuracy compared to existing LiDAR-only methods while maintaining a real-time processing rate.