SAMFNet: Scene-aware sampling and multi-stage fusion for multimodal 3D object detection

Recently, multimodal 3D object detection (M3OD) that fuses the complementary information from LiDAR data and RGB images has gained significant attention. However, the inherent structural differences between point clouds and images pose fusion challenges, significantly hindering the exploration of correlations within multimodal data. To address this issue, this paper introduces an enhanced multimodal 3D object detection framework (SAMFNet), which leverages virtual point clouds generated from depth completion. Specifically, we design a scene-aware sampling module (SASM) that employs tailored sampling strategies for different bins based on the density distribution of point clouds. This effectively alleviates the detection bias problem while ensuring the key information of virtual points, significantly reducing the computational cost. In addition, we introduce a multi-stage feature fusion module (MSFFM) that embeds point-level and regional-adaptive feature fusion strategies to generate more informative multimodal features by fusing features with different granularities. To further improve the accuracy of model detection, we also introduce a confidence prediction branch unit (CPBU), which improves the detection accuracy by predicting the confidence of feature classification in the intermediate stage. Extensive experiments on the challenging KITTI dataset demonstrate the validity of our model.