cs.CVAug 16, 2026

YOLO26-RD: An End-to-End Road Damage Detection Network With Learnable Contrast Enhancement and Edge-Guided Downsampling

Authors: Sompote Youwai, Pawarotorn Chaipetch, Hathairat Samaikul, Theerayut Yonseng

Organizations: AI Research Group, Department of Civil Engineering, King Mongkut’s University of Technology Thonburi, Bangkok, Thailand · Infraplus Co., Ltd., Thailand

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Pavement distress detectors are conventionally specialised for small objects, typically by adding a stride-4 detection head and replacing strided convolution with space-to-depth downsampling. This paper tests that premise against the annotation geometry of region level survey imagery and finds it fails: 1.28% of instances are small at 640 resolution while 70.37% are large, yet a stride-4 level would claim 75.3% of anchors, and complete misses rather than localisation errors dominate baseline failures. YOLO26-RD therefore reallocates the anchor budget, retaining the stride-4 branch as neck features but carrying no detection level there, and adds LearnableContrast, a 494 parameter per tile correction learned from the detection loss and active at inference, and EdgeSPD, a lossless space-to-depth downsampler gated by a fixed Sobel prior. Fifteen models were trained from scratch under one recipe, five scales each of YOLO26-RD and of matched YOLO26 and YOLOv12 families. Averaged over scales YOLO26-RD returns 0.790 mAP50 and 0.482 mAP50-95 against 0.776 and 0.471 for YOLO26 and 0.755 and 0.468 for YOLOv12; it exceeds both on mAP50 at every scale from s upward, and at m, l and x it leads on both metrics, twelve pairwise comparisons decided without exception. YOLO26-RD-l is the best of the fifteen at 0.809 mAP50 and 0.497 mAP50-95, improving on the YOLO26 reference by 0.031 and 0.030 and leading all six per class entries; every arm of a module ablation also exceeds that reference. The margin is thus a property of the architecture rather than of one tuned configuration, though three of the twelve margins lie inside the dataset 0.015 resolution limit and the held out split reproduces the ordering against YOLO26 but not YOLOv12 at scale x. As a TensorRT FP16 engine the released model sustains 98 frames per second on an entry level accelerator, against the 21 needed at 100 km/h.

Figures & tables

Fig. 1: Concept of the automated survey. Frames are captured continuously, processed on board as they arrive, and reduced to one structured record each; only those records are transmitted, while the imagery is retained in the vehicle. The detector must therefore sustain the capture rate on hardware carried in the vehicle. Quantities are given in Section 8.
Method (venue, yr)BaseKey additionsContrastDownsamplingP2 headDataset — reported Δ mAP50
YOLOv8-PD (Sci. Rep. 2024)YOLOv8nBOT transformer, LSKA attention, C2fGhost, shared light head—strided convnoRDD2022 — 0.706 (+1.4)
EMG-YOLO (Front. Neurorob. 2024)YOLOv5global-context C3, decoupled head, MPDIoU—strided convnoRDD2022 — (+2.9); 2× model size
SRD-YOLO (AICIT 2024)YOLOv8RFAConv, 4-level scale fusion for small targets—strided convyespavement distress — (+2.7)
MFF-YOLO (NNICE 2024)YOLOv8BoTNet backbone, BiFPN neck, small-target focus—strided conv—road crack, complex background
SDFC-YOLO [14]YOLOdynamic downsampling module, feature-fusion compensation—adaptive sampling positions—pavement distress
YOLOv11n-CDL (JCEM 2026)YOLO11ndouble-stage attention, P2 small-object path—strided convyesfine cracks / micro-potholes
TABLE I: Related work summary. Contrast, downsampling and P2 give each method’s treatment of low-contrast input, early feature decimation, and whether a stride-4 detection head is added.
Fig. 2: Survey platform: (a) the instrumented vehicle, and (b) the roof-mounted instrument rail carrying the camera housing.
splitimagesbackgroundsalligatorcrackpatching
train6,5631,711 (26%)1,7425,380662
test50439 (8%)24141865
val551117 (21%)23738564
TABLE II: Dataset splits: image counts, background frames, and per-class instance counts.
Fig. 3: Down-facing survey frames with YOLO26-RD detections overlaid, annotated by class, confidence and physical extent. (a) An alligator crack region (red box). (b) A crack under dappled tree shadow (green box). (c) A crack and a patching region in one frame (green and blue boxes). Discussed in the text above.
Fig. 4: YOLO26-RD topology. (a) End-to-end data flow, blocks annotated with output channels and stride; orange marks the two proposed modules, green the area-attention blocks, and dashed green arcs the BiFPN-style backbone skips. (b) Internals of LearnableContrast and EdgeSPD. Specified layer by layer in Table A2 and discussed in Sections 4.1 to 4.3.
Fig. 5: LearnableContrast data flow. The identity path (top) carries the image unchanged to the output rule, while the correction path (bottom) reduces it to per-tile statistics, predicts the per-tile corrections with the module’s only learnable block (orange, 494 parameters) and upsamples them before the pixelwise application. Blue blocks have no learnable parameters. Discussed in the text above.
Fig. 6: EdgeSPD data flow. The edge-gate path (bottom) computes a Sobel gradient-magnitude map from the channel mean, with no learnable parameters, and maps it through the module’s entire learnable budget of 2 parameters (orange) into a multiplicative gate. The gate amplifies edge-bearing locations before the lossless space-to-depth rearrangement and 3×3 fusion (top right). Discussed in the text above.
itemsetting
Hardware1 × NVIDIA RTX 5090 (32 GB); 2 × RTX 5090 for the scale-x arms, which do not fit one card at batch 32
FrameworkPyTorch 2.12 / CUDA 13; Ultralytics 8.4.115 fork containing both modules
Initializationfrom scratch; no pretrained weights in any arm
Input resolution640 × 640
Batch32, with gradient accumulation to a nominal batch of 64
Epochs150
TABLE III: Training and evaluation configuration, identical for every arm.
 baseline: YOLO26-lYOLO26-RD-labsoluterelative
mAP500.7780.809+0.031+3.98%
mAP50-950.4670.497+0.030+6.42%
precision0.8570.830−0.027−3.15%
recall0.7060.717+0.011+1.56%
AP50, alligator0.8210.854+0.033+4.02%
AP50, crack0.7210.744+0.023+3.19%
TABLE IV: YOLO26-RD-l against the YOLO26-l reference on the test split.
#armLCEdgeSPDmAP50mAP50-95PRAP50 alligatorAP50 crackAP50 patching
—YOLO26-l baseline (not the RD graph)no—0.7780.4670.8570.7060.8210.7210.792
B1LearnableContrast onlyyes00.7820.4870.7870.7440.8050.7370.804
B2EdgeSPD, one locationno10.7900.4810.8380.7140.8390.7420.790
B3EdgeSPD, two locationsno20.7940.4940.8540.6840.8420.7350.806
B4LearnableContrast + EdgeSPD, one locationyes10.7860.4930.8130.7080.8450.7270.787
—YOLO26-RD-l (proposed, Table 4)yes20.8090.4970.8300.7170.8540.7440.828
TABLE V: Module ablation of YOLO26-RD-l on the 504-image test split. LC is the LearnableContrast stem; EdgeSPD gives the number of gated downsampling locations, disabled locations being replaced by parameter-matched plain space-to-depth. The baseline row is the unmodified YOLO26-l of Table 4 and is outside the proposed graph. No arm was evaluated on the validation split (Section 7.1). Bold marks the best value per column.
modelparamsmAP50mAP50-95
YOLO26-RD (proposed)
YOLO26-RD-n3.13M0.7570.460
YOLO26-RD-s12.42M0.7870.469
YOLO26-RD-m31.61M0.7940.488
YOLO26-RD-l (Table 4)36.22M0.8090.497
YOLO26-RD-x81.36M0.8010.495
TABLE VI: Benchmark comparison across scale ranges. All fifteen arms trained from scratch under the 150-epoch recipe of Table 3 on the same dataset, splits and resolution, each evaluated at its fitness-selected checkpoint on the 504-image test split. Bold marks the best value per column within each family.
Fig. 7: Scale comparison of YOLO26 (blue circles), YOLO26-RD (orange squares) and YOLOv12 (green triangles, dashed) against parameter count on a logarithmic axis, scales n/s/m/l/x per family, at the fitness-selected best.pt checkpoint of every run. (a) mAP50. (b) mAP50-95. All points are selection-split values at completed runs. Values are tabulated in Table 6.
Fig. 8: Accuracy against TensorRT FP16 inference latency on an RTX 5090, at scales n/s/m/l/x: YOLO26 (blue circles), YOLO26-RD (orange squares), YOLOv12 (green triangles, dashed). (a) mAP50. (b) mAP50-95. All points are selection-split values, as in Figure 7; protocol and excluded stages are given in the text above.
Fig. 9: Throughput of YOLO26-RD across deployment hardware, as TensorRT FP16 engines at batch one and 640² input. (a) Measured throughput by scale on three devices, each bar labelled with its value; the dashed line marks the 20.6 frames per second that a 100 km/h survey requires at half-frame overlap (Section 8). (b) The same throughput against selection-split mAP50, with the five scales labelled on the RTX 5060 series; all three series carry the same checkpoints, so the curves differ from one another only by a horizontal shift. Latency in milliseconds is the reciprocal of the plotted throughput. The RTX 5090 and RTX 5060 series time the inference stage alone and the RTX 5070 series is end-to-end, and the three were measured under different benchmarking harnesses, so they are not quantitatively comparable to one another; see the text above. Parameter counts and the accompanying mAP50-95 figures are given in Table 6.
Fig. 10: Channel-mean activation maps for YOLO26-l (upper row of each panel) and YOLO26-RD-l (lower row) on three selection-split frames: (a) a patching region, (b) two full-height longitudinal cracks, (c) an alligator-cracked area. Columns, stage indices and colour scale are specified in the text above.
Fig. 11: Four representative failures of YOLO26-RD-l on the selection split: (a) a small boundary-truncated alligator region missed beside a correctly detected larger one; (b) a faint isolated crack with no prediction; (c) three elongated or boundary-clipped patching regions with no prediction; (d) a dense seven-crack scene. Ground-truth boxes are coloured by outcome, green for a hit, amber for an extent error and red for a miss, and predictions are cyan. Selection procedure and matching thresholds are given in the text above.
audit stepmeasurementimplication
Instance size (640² input)1.28% small, 28.34% medium, 70.37% large; smallest instance √area 14.4 px; small share 0.02% at 1024² and 0.01% at 1280²the small-object premise fails at every resolution considered
Annotated box geometrymedian crack box 138 × 1,476 px at aspect 9.5:1 (p95 17.2:1), spanning about 90% of frame height; alligator median √area 1,020 px; patching median aspect 0.98instances are thin, not small, and no class occupies a small-object regime
Anchor allocation (640² input)P3, P4 and P5 supply 6,400, 1,600 and 400 anchors, 8,400 in total; a stride-4 level adds 25,600, or 75.3% of all anchorsa stride-4 detection level is not warranted on this data
Failure decomposition (YOLO26-l, 724 instances)missed entirely 25 of 241 alligator (10%), 67 of 418 crack (16%), 22 of 65 patching (34%); localization error 30 of 418 crack (7%) and 0 of 65 patching; misclassified 2 in the whole splitthe deficit is sensitivity, not box regression
Geometry shift, train to test (crack)median crack spans 0.90 of frame height in training against 0.51 in the test split; transverse share 18.9% against 38.3%part of the crack miss rate is a geometry difference between the splits, not a capacity limit
Label completeness (confidence ≥ 0.5, IoU ¡ 0.1)the majority of confident unmatched detections are real but unlabelled distress; 74 of the 113 false positives at IoU ¡ 0.3 belong to crackmeasured precision is a lower bound, depressed equally for every arm
TABLE VII: Data audit of the imagery and of the unmodified YOLO26-l reference. Size classes follow the COCO convention applied after letterboxing. The failure decomposition is exhaustive per ground-truth instance under one-to-one matching over the 724 test-split instances, at an operating point returning precision 0.828, recall 0.675, mAP50 0.704 and mAP50-95 0.421, which is not the framework’s default and is therefore not interchangeable with the dataset-level metrics reported elsewhere.
#frommodulenc_outparamsstriderole
0−1LearnableContrast13494/1tile-wise contrast stem (proposed)
1−1Conv 3×3 s21641,856/2P1
2−1Conv 3×3 s2112873,984/4P2
3−1C3k22256173,824/4 
4−1EdgeSPD12562,359,810/8P3 edge-gated lossless downsample (proposed)
5−1C3k22512691,712/8backbone P3 out
TABLE VIII: YOLO26-RD layer-by-layer specification at scale l (depth 1.00, width 1.00, maximum 512 channels). Parameter counts are per layer for the unfused model at nc = 3 . Groups of three layers are given as one row with their summed count, the n and c_out entries being those of the fusion block terminating the group.
More in this topic

Object Detection

Object detection locates and classifies objects in images or video, balancing localization accuracy, robustness, and inference speed.

See all 624 papers

Explore similar work

CardsList
Aug 9, 2026

RMR-Net: Degradation-Evidence-Guided Road-Image Restoration for Defect Detection

Vehicle-mounted road cameras are vulnerable to motion blur, defocus, poor illumination, and noise, which can erase thin cracks and pothole boundaries needed by road defect detectors. This paper presents RMR-Net, a compact task-aware restoration front end that estimates degradation evidence from the image, optionally fuses it with existing corruption context/parameters, conditions lightweight restoration blocks, and returns high-frequency pavement detail through a bounded residual path. The experimental scope is deliberately controlled: the conditioning information used on the Image and Vision Computing New Zealand (IVCNZ) pothole dataset and the Road Damage Dataset: Potholes, Cracks and Manholes (PCM) consists of saved synthetic-generator parameters, not measured vehicle telemetry. A clean-trained, frozen YOLO11s detector evaluates every image source. Across eight held-out degradation conditions, RMR-Net obtains the highest mAP50 in seven, including 0.140-0.427 for IVCNZ motion blur and 0.060-0.233 for PCM defocus. A compact ablation identifies the bounded detail path as the largest local contributor, while degradation conditioning and detector-aware stability terms provide complementary guidance.

Amir Ghorbani, Amirali K. Gostar, WeiQin Chuah +3

School of Engineering, RMIT University, Melbourne, VIC 3000, Australia

Image Corruption RobustnessImage RestorationDegradation-Aware Image Restoration
AskPDF
Aug 27, 2026

Depth-Aware Pothole Detection Using YOLO and RT-DETR at the Edge

Pothole detection and its severity measurement is still an important challenges in urban infrastructure management, where late maintenance directly contributes to vehicle damage, road accidents, and escalating repair costs. Existing automated approaches depend on 2D RGB images and cannot measure physical depth of potholes. In this paper, we present a depthaware pothole detection framework and then compare five architectures: YOLOv8n, YOLOv8nSeg, YOLOv9t, RTDETRL, and RTDETRX for RGB-D sensor fusion-based detection and automated depth measurement. A custom offline augmentation pipeline is used here to simulate adverse road monitoring conditions. All models are trained on the PothRGBD dataset with an 80% training and 20% validation split and evaluated using Precision, Recall, mAP@50, and mAP@50_95. Before measuring the depth data, all depth maps are corrected for camera tilt using RANSAC ground-plane orthorectification and all zero-valued sensor pixels are cast to NaN before any statistic is computed. YOLOv8nSeg achieves the highest mAP@50 of 0.9556 and mAP@50_95 of 0.6758 with the most accurate depth estimate of 2.96 cm with the pixel-precise Dseg algorithm. YOLOv8n achieves the fastest inference at 3.6ms. RTDETRX achieves the highest detection confidence at 92.70%. An important finding is that even after full RANSAC orthorectification, bounding box models overestimate pothole depth by 0.16 to 0.21 cm compared to pixel precise segmentation masks. This confirms that the pavement inclusion bias is structural rather than a calibration artifact.

Md Monjurul Ahsan Prodhan, Md Nour Hossain

College of Emergency Preparedness, Homeland Security and Cybersecurity University at Albany, State University of New York Albany, United States

Object DetectionDepth Estimation
AskPDF
Jun 2, 2026

Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models

Real-time vision demands models that are accurate, efficient, and simple to deploy across diverse hardware. The YOLO family has become widely deployed for this reason, yet most YOLO detectors still rely on non-maximum suppression at inference, carry heavy detection heads due to Distribution Focal Loss, require long training schedules, and can leave the smallest objects without positive label assignments. We present Ultralytics YOLO26, a unified real-time vision model family that addresses these limitations through coordinated architecture and training advances. YOLO26 uses a dual-head design for native NMS-free end-to-end inference and removes DFL entirely, yielding a lighter head with unconstrained regression range. Its training pipeline combines MuSGD, a hybrid Muon-SGD optimizer adapted from large language model training; Progressive Loss, which shifts supervision toward the inference-time head; and STAL, a label assignment strategy that guarantees positive coverage for small objects. Beyond detection, YOLO26 introduces task-specific head and loss designs for instance segmentation, pose estimation, and oriented detection, producing consistent gains across tasks and scales. The family spans five scales (n/s/m/l/x) and supports detection, instance segmentation, pose estimation, classification, and oriented detection in a single pipeline, with an open-vocabulary extension, YOLOE-26, for text-, visual-, and prompt-free inference. Across all scales, YOLO26 achieves 40.9-57.5 mAP on COCO at 1.7-11.8 ms T4 TensorRT latency, advancing the accuracy-latency Pareto front over prior real-time detectors, while YOLOE-26x reaches 40.6 AP on LVIS minival under text prompting. Code and models are available at https://github.com/ultralytics/ultralytics.

Glenn Jocher, Jing Qiu, Mengyu Liu +3

Ultralytics

Open-Vocabulary Object DetectionInstance SegmentationEfficient Object DetectionObject DetectionReal-Time Object Detection
AskPDF