BMGKD: A High Precision Object Detection Knowledge Distillation Method for Bridging Multi-Dimensional Gaps

Existing knowledge distillation methods for object detection struggle to bridge the teacher-student capacity gap and overlook the inherent differences between classification and regression subtasks. To address these issues, we propose a Bridging Multi-dimensional Gaps Knowledge Distillation (BMGKD) method, which comprises two core modules: a feature difference distillation module and a response difference distillation module. The feature difference distillation module achieves global feature structural alignment via improved centered kernel alignment and performs local key feature alignment using joint spatial and channel-wise cosine similarity masks. The response difference distillation module constructs a dynamic classification mask and a high-quality prediction box selection mechanism, along with a classification and regression co-optimization loss function. On the MS COCO dataset, BMGKD achieves the highest mAP across all three teacher-student configurations on two-stage Faster R-CNN, single-stage anchor-free GFL, and single-stage anchor-based RetinaNet detectors. In the ResNet101 $\to$ResNet50 setting, it surpasses baselines by 2.7, 4.2, and 2.8 percentage points, respectively, and maintains consistent optimality under varying capacity gaps and cross-architecture scenarios. On Pascal VOC 2007, BMGKD attains mAP of 56.2 $\mathrm{\%}$, 58.3 $\mathrm{\%}$, and 58.0 $\mathrm{\%}$ on the same three detectors, outperforming all compared methods. Ablation studies confirm each component’s contribution. These results demonstrate that BMGKD effectively resolves the distillation accuracy degradation caused by capacity gaps and task discrepancies between teacher and student models. It provides an effective solution for knowledge distillation across diverse detection architectures.