High-Order Structural Relation Distillation Networks From LiDAR to Monocular Image 3D Detectors

3D object detection is a crucial and complex undertaking in the realm of 3D scene comprehension. Monocular-based 3D detectors, in comparison to LiDAR 3D detectors that utilize point clouds as input, often exhibit a significant performance gap. Incorporating guidance from LiDAR-based detectors has led to notable advancements in monocular 3D detection. Nevertheless, some current approaches focus solely on transferring feature or response knowledge from LiDAR data, neglecting the valuable geometric structural information present in the LiDAR data. As a result, there is untapped potential to fully exploit the geometric relationships within the LiDAR data to further improve monocular 3D detection performance. In this paper, we propose a High-order Structural Relation Distillation Networks, which establishes the geometric structure information in the BEV feature level and distills the relation from LiDAR Detectors to Monocular Image Detectors. Specifically, to better align the structural relation with the LiDAR BEV features, we propose the Image BEV Feature Enhancing Module, which can capture essential spatial depth context via explicitly modeling the interdependence between the depth channels of BEV features to improve the representations. In addition, instead of only distillation at feature level, our distillation model contains point, line, and surface relation distillation, which captures high-order structural relations from LiDAR BEV features and enhances the detection of the image-based student model, particularly for non-rigid object detection. Our method achieves state-of-the-art performance, especially in the case of non-rigid Cyclist class.