A Appendix

A.1 Performance breakdown for categories In table 1, we report the performance on each category. It is shown that our DeepInteraction performs the best among all the competitors across the most of object categories. Table 1: Comparison with state-of-the-art methods on the nuScenes test set. 'L' and'C' represent LiDAR and camera, respectively. A.2 Discussions of potential societal impacts Fusing multi-modal information allows to compensate for the shortcomings of single modality in 3D object detection, leading to more more accurate and robust performance.