StereoDistill: Pick the Cream from LiDAR for Distilling Stereo-Based 3D Object Detection

In this paper, we propose a cross-modal distillation method named StereoDistill to narrow the gap between stereo and LiDAR-based approaches via distilling detectors from superior LiDAR model at response level, which is usually overlooked in 3D object detection distillation. The key designs of are: X-component Guided Distillation~(XGD) for regression Cross-anchor Logit Distillation~(CLD) classification. XGD, instead empirically adopting threshold select high-quality teacher predictions as soft targets, decompose predicted box into sub-components retain corresponding part if component pilot consistent with ground truth largely boost number positive alleviate mimicking difficulty student model. For CLD, aggregate probability distribution all anchors same position encourage highest anchor rather than individually distill level. Finally, our achieves state-of-the-art results stereo-based on KITTI test benchmark extensive experiments Argoverse Dataset validate effectiveness.