MLF3D: Multi-Level Fusion for Multi-Modal 3D Object Detection

Recently, 3D object detection techniques based on the fusion of camera and LiDAR sensor modalities have received much attention due to their complementary capabilities. How-ever, prevalent multi-modal models are relatively homogeneous in terms of feature fusion strategies, making their performance being strictly limited to the detection results of one of the modalities. While the latest data-level fusion models based on virtual point clouds do not make further use of image features, resulting in a large amount of noise in depth estimation. To address the above issues, this paper integrates the advantages of data-level and feature-level sensor fusion, and proposes MLF3D, a 3D object detection based on multi-level fusion. MLF3D generates virtual point clouds to realize the data-level fusion, and implements feature-level fusion through two key designs: VIConv3D and ASFA. VIConv3D reduces the noise problem and realizes deep interactive enhancement of features through cross-modal fusion, noise sensing, and cross-space fusion. ASFA refines the bounding box by adaptively fusing cross-layer spatial semantic information. Our MLF3D achieves 92.91%, 87.71% AP and 85.25% AP in easy, medium and hard scenarios on the KITTI's 3D Car Detection Leaderboard, realizing excellent performance.