YOLO-GRAT: A Gated Residual Cross-Modal Fusion Network with Adaptive Triplet Attention for Visible–Infrared Object Detection

YOLO-GRAT integrates P3 Cross-Modal Gated Residual Fusion with adaptively weighted Triplet Attention inside backbone C3 bottlenecks of a YOLOv5-based visible–infrared detector. The fusion block combines input-dependent modality mixing, a mean-residual path and pooled-token Transformer refinement; the attention extension learns aggregation weights for three established rotated branches. In the FLIR comparison, each module is added independently to a common dual-stream baseline. Full gives 83.4/44.5% AP50/AP50:95 versus 80.7/42.1% for that baseline. A separate comparison uses four identified fixed-final M3FD seed-0 checkpoints on the same 840 validation pairs. Their official COCO AP is 57.21% for CFT, 57.48% for CGRF only, 57.40% for weighted Triplet only and 56.85% for Full; Full small-object AP is 35.46 versus CFT 35.95%. The FLIR configuration results motivate the combination, while the identified M3FD results reveal metric-dependent trade-offs. The study evaluates complementary within-stream recalibration and cross-modal interaction at the configuration level, with conclusions scoped to the reported datasets, metrics and training records.