A Fully Sparse Detector with Cascade Heads for LiDAR-based 3D Object Detection

LiDAR-based 3D object detection is a critical component of perception systems in autonomous driving. Existing detectors typically rely on dense feature maps for 3D object prediction. However, as the perception range expands, the computational overhead associated with dense feature maps increases quadratically, creating significant challenges for scaling to long-range detection scenarios. To address this issue, fully sparse detectors have recently been proposed, but they often suffer from suboptimal performance or excessively complex architectures. In this paper, we present CHNet, a fully sparse detector designed to overcome these limitations by introducing a cascade head module, which consists of two sequential heads. By generating coarse object predictions, creating new feature points based on centers of the predicted objects, and aggregating features of the entire object to these new center points, CHNet enhances the qualities of object feature representations, achieving higher detection performance. Notably, CHNet effectively addresses the center feature missing problem in LiDAR-based 3D object detection by adding at most a hundred new feature points, offering a potential avenue for future research. On the widely used Waymo Open and Argoverse2 datasets, CHNet achieves impressive results without any additional embellishments.