Abstract
Automated pavement-distress detection is commonly framed as a small-object problem, motivating high-resolution P2/4 detection heads and lossless downsampling. We present YOLO26-RD, an end-to-end (NMS-free) detector built on YOLO26 with two lightweight novel modules (LearnableContrast, a 494-parameter differentiable analogue of CLAHE that adapts contrast per tile inside the network, and EdgeSPD, a Sobel-gated space-to-depth downsampler adding only 2 parameters over SPD-Conv), and we subject the design to a data-first audit on a 7,618-image road-survey dataset (alligator crack, linear crack, patching). The audit falsifies the small-object premise: 92% of instances are COCO-large, and linear cracks are extreme-aspect structures (median 10:1) whose difficulty is sensitivity, not localization. Guided by this analysis, we remove the P2 detection level while retaining P2 features in the fusion path, which improves mAP50 by 2.8 points over the full YOLO26-RD model and reduces epoch time by 8%. Trained from scratch at 640x640, our best screening configuration reaches 0.787 mAP50 on the validation split versus a 0.771 project baseline (a stock YOLO26-s of uncontrolled recipe), with the largest per-class gain on the rarest class (patching, 2.6 points over baseline; 9.4 over the unmodified YOLO26-RD control under an identical recipe). A failure-mode decomposition further attributes the residual error of the bottleneck class (crack, approximately 0.74 across all architectures tested) to train/validation distribution shift on crack orientation and length, sub-pixel crack width at 640x640, and label incompleteness, factors no architecture change can address. We argue that for pavement imagery, measurement-driven subtraction outperforms module accretion, and we release our audit protocol alongside the model.
Abstract (translated)
URL
https://arxiv.org/abs/2608.15713
