CoKNet: Colorized key-instance guided fusion for 3D object detection
Xiangzhen Chang, Zhuochen Yu, David K. Y. Yau, Liping Yan, Yuanqing Xia · Pattern Recognition Letters · 2026
Fusing LiDAR with virtual points generated by depth completion from RGB images and LiDAR has emerged as a promising direction for 3D object detection. However, it remains challenging to balance dense geometric coverage, reliable instance-level semantics, and robustness to noisy virtual points. To address these issues, we propose CoKNet , a colorized key-instance guided fusion network. CoKNet consists of two key designs: KICP (Key Instance Guided Color Painting), which establishes instance-level anchors by enriching LiDAR points with depth-consistent color information from CKIs, and KIGF (Key Instance Guided Fusion), which leverages these anchors to achieve complementary perception between instance and global features through progressive BEV–RoI fusion. Experiments on the KITTI and nuScenes benchmarks demonstrate that CoKNet improves 3D detection performance over recent LiDAR–camera fusion methods, particularly in long-range, occluded, and image-degraded scenarios.