A Coarse-to-Fine Cascade for Fine-Grained Small-Object Traffic Sign Detection: Revisiting the Category-Granularity Bottleneck

Traffic signs in high-resolution street-view images are small and belong to many fine-grained classes, where one-stage detectors perform poorly. A factorial analysis shows that the dominant bottleneck is fine-grained inter-class confusion rather than feature resolution: on an identical YOLO26n, adding a P2 head to detect all 45 classes slightly lowers the mean average precision at an IoU threshold of 0.5 (mAP@0.5) from 73.1% to 72.7%, whereas merging the 45 classes into three semantic coarse categories (mandatory, prohibitory, warning) raises it to 89.7%. We therefore propose a coarse-to-fine cascade: a lightweight YOLO26n with an added P2 layer first localizes the three coarse categories, and equal-size 128 × 128 crops are then classified into 45 classes by a YOLO11n classifier; group masking, score multiplication and confidence-adaptive fine classification enforce semantic consistency, calibrate scores and allocate computation on demand. On TT100K, the cascade reaches 89.5% mAP@0.5 for 45 classes, 16.4 points above the direct baseline, with 7.0–10.8-point gains across object size bins, using only 4.11 M parameters and 3.77 GFLOPs at about 39 FPS. In training-free cross-domain tests on CCTSDB2021 and GTSDB, coarse-class merging brings 5.8 and 10.0 points, respectively, and adaptive fine classification gives the best accuracy in each domain.