Controlled Reverse-Order Gradient Exchange in Bidirectional Feature Fusion

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

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. In RT-DETR-R18 experiments on VisDrone2019-DET, with the original P4–P5 path retained, reverse-order exchange improves the 30-epoch mean mAP50 by 0.614 and mAP50-95 by 0.438 percentage points over independent kernels, and outperforms same-order exchange and two reception-control ablations. Under a common COCO-style reevaluation, the gains concentrate on small objects, whose AP rises by 0.779 percentage points, accompanied by 167 additional correct matches and reductions in duplicate, wrong-class, localization, and background errors. In the studied P3–P4 configuration, controlled reverse-order exchange thus turns an existing structural connection into useful training cooperation, improving small object detection in UAV imagery while preserving independent learning at both endpoints and leaving inference unchanged.