Controlled Reverse-Order Gradient Exchange in Bidirectional Feature Fusion

Bidirectional P3–P4 fusion continuously connects high-resolution features with semantic information from the adjacent level, providing cross-level information for small object detection. The channel kernels in the two directions adapt to their respective inputs, while this feature connection also provides a basis for cooperative learning. We propose controlled reverse-order gradient exchange to exploit the learning signal from the opposite endpoint while retaining independent kernels. Reverse-order mapping pairs opposite offsets in local channel interactions. Each receiver selects directions compatible with its own gradient and limits the auxiliary contribution relative to that norm. The mechanism operates only during training and leaves the inference path unchanged. On VisDrone2019-DET with RT-DETR-R18, reverse-order exchange yields mean gains of 0.419 mAP50 and 0.254 mAP50–95 percentage points over independent kernels across three training seeds. Mean small-object AP increases by 0.323 points. At confidence 0.25, prediction matching yields 105.3 additional correct matches on average. These results suggest that controlled cooperation between independent channel kernels can benefit small object detection.