Making radar better at seeing: why simpler fusion beats complexity

3D object detection for self-driving cars relies on fusing radar and camera data, but sparse radar points have always been a bottleneck. This work flips the conventional wisdom: instead of building complex fusion modules, the authors show that improving how you extract features from radar alone achieves better accuracy faster. RCGDet3D uses ray-aligned coordinate prediction to make geometric learning easier, then adds visual cues from cameras afterward. On two benchmark datasets, it beats prior methods in both speed and accuracy—with clear potential for real-world deployment.