Consistent Feature Alignment for Cross-Modal Knowledge Distillation in Monocular 3D Object Detection

Cross-modal knowledge distillation (CMKD) in monocular 3D object detection transfers LiDAR's accurate depth information to compensate for the limitations of camera model. However, current methods directly align the intermediate features of the teacher and student networks, in which the modality gap between LiDAR and camera hinders their effectiveness. To mitigate this issue, we design two modules, namely, Consistent Alignment Module (CAM) and Deformable Adapter Module (DAM) to reduce the modality gap of CMKD. The CAM transforms intermediate features of LiDAR and camera into some consistent features through a lightweight Target Head. It is based on the observation that some high-level features such as heatmaps and depths are highly correlated in CMKD, though modality gap appears between LiDAR and camera. Therefore, these features can be effectively transferred from teacher to student in CMKD. The DAM introduces a deformable adapter for the intermediate features of the student network to reduce background noise in CMKD. This helps to dynamically align its intermediate features with the teacher network. We then propose a Consistent Feature Alignment network (MonoCFA) for CMKD to boost monocular 3D object detection. Our network integrates the two designed modules at different levels of the teacher and student networks, in order to align the intermediate features of LiDAR and camera more accurately and reliably. Our model can be widely applied to existing monocular 3D object detection models. For validation, we choose the representative MonoDLE, GUPNet, and DID-M3D as base models. Experiments on the KITTI benchmark show that our method significantly outperforms the three base models by 39%, 15.5%, and 15%, respectively, and achieves state-of-the-art when compared to other CMKD models.