A LiDAR-camera 3D object-detection network that unites point-level and feature-level fusion by exploiting the geometric hard association from sensor calibration matrices rather than relying solely on a learned soft association. In early fusion, point clouds are decorated with 2D CNN image features and encoded by independent sparse convolutions; in mid-level fusion, detection queries are initialized from a center heatmap with embedded predicted class labels so query positions start near true object centers. Using calibration-based hard association to guide query generation improves interpretability and accuracy, outperforming prior fusion methods on the KITTI and Waymo benchmarks.
This paper presents DecoratingFusion, a LiDAR-camera fusion network for 3D object detection in autonomous driving that combines point-level and feature-level fusion. State-of-the-art fusion works at …