BR-AsPavDam - Brazilian Asphalt Pavement Damage Dataset

BR-AsPavDam: A public urban asphalt-defects dataset collected in Brazil with a bumper-mounted smartphone. Images include varied lighting, traffic and surface conditions. Five classes follow DNIT guidelines: Fissures, Shoving (rutting/shoving), Ravelling, Pothole, Patch. Labels were curated by a technical committee with quality checks. YOLO-format annotations (one TXT per image, class x_center y_center width height, normalized). High-resolution image dataset of defects in urban asphalt pavement in Brazil, annotated in YOLO format. Each image is accompanied by a .txt file with normalized bounding boxes (cx cy w h). Classes 0. Fissures Shoving Ravelling Pothole Patch Structure: images/, labels/, splits train/val/test. Resolution: high-res originals; filenames are stable; checksums provided. Intended use: benchmarking defect detection and public-works planning. Known limitations: class imbalance, weather and location bias. Related resources: Source code & training recipes (GitHub), thesis with methods and metrics (mAP@50 = 0.844 for best model). License: CC-BY 4.0 (data). Please cite the dataset and repository; BibTeX provided below. How to reproduce: see GitHub for enviroment, scripts, and release tags matching this version. @dataset{Borges_BR-AsPavDam_2025, title = {BR-AsPavDam: Urban asphalt defects dataset (Brazil)}, author = {Borges, Marcos A. and Spanhol, Fabio}, year = {2025}, publisher = {Zenodo}, doi = {10.5281/zenodo.16291115}, url = {https://doi.org/10.5281/zenodo.16291115}}____________More details: Field Data Acquisition The data were collected using a video-capture device (Apple iPhone 15) mounted on the front bumper of a passenger vehicle that drove along urban roads in Cascavel, Paraná, Brazil. Recording was started and stopped manually on the device. Whenever possible, the vehicle speed was kept at 40 km/h while maintaining a safe distance from the vehicle ahead. Recording settings (native Camera app): Resolution: 4K (3840 × 2160 px), capturing fine-grained surface details (e.g., fissures and irregularities) Frame rate: 60 FPS, to ensure fluid motion appropriate for frame extraction Exposure & lighting: natural daylight, spanning sunny and shaded segments Aperture / ISO / Focus: f/1.8, ISO 100, continuous autofocus, optical image stabilization, auto white balance Data collection took place between 09:30 and 14:30, to ensure adequate illumination. Routes were chosen to include clear occurrences of asphalt distresses, maximizing the likelihood of capturing diverse defect types within the available driving time. Surveyed road segments (urban network of Cascavel-PR): Segment 1: Rua Vitória, Rua Barão do Cerro Azul, Rua São Paulo — ~3.04 km Segment 2: Rua Rio Grande do Sul, Rua Erechim — ~3.44 km Segment 3: Rua Paraná — ~2.10 km Segment 4: Rua Natal — ~1.04 km Segment 5: Rua João Lili Cirico, Rua Teresina, Rua Rui Barbosa, Rua São Luís, Rua Visconde do Rio Branco — ~3.61 km The total route length was approximately 13.22 km, spanning central arteries with higher flows and heavier vehicles (Segments 1–3) and residential/local streets with two-way traffic and lighter flow (Segments 4–5). The segments were selected by municipal civil engineers based on prior knowledge of problematic locations, ensuring that all normative defect types targeted in this study were represented despite the modest overall distance. Quality Control (Frame Extraction) After field acquisition, video files were transferred to an Apple MacBook (M2). Static images were extracted using VLC(frame capture), preserving the original quality, and saved as PNG. To improve visibility and variety, frames with very near vehicles, traffic lights stops, or heavy congestion were avoided. To reduce redundancy, screen captures were taken at ~2-second intervals; combined with the target speed (≈ 40 km/h), this spacing limits near-duplicate imagery. A total of 3,565 images were initially extracted; not all were retained, as some contained no defects. Annotation Annotations were produced with CVAT by specialists in asphalt pavements. The labeling committee included three civil engineers from the Cascavel municipality, all with prior experience in paving projects and knowledge of the Brazilian DNIT standards. Candidate engineers were interviewed regarding practical paving experience and familiarity with the normative criteria. Bounding-box annotations followed DNIT’s defect definitions. During pilot labeling, eight DNIT categories were initially targeted: cracking (fissures), rutting (afundamento), corrugation (ondulação), shoving (escorregamento), bleeding (exsudação), ravelling (desgaste), pothole (buraco), and patching (remendo). Two adjustments were made to improve reliability: Bleeding (exsudação) was removed due to frequent confusion with standing water from rainfall, washing, or leaks. The fully automated diagnosis of this phenomenon would likely require external/auxiliary informationbeyond RGB imagery. Corrugation, shoving, and rutting were merged into a single Shovingclass, given their similar plastic-deformation mechanism and visual overlap under DNIT criteria, and their low frequency in urban settings(heavier-vehicle loads are more common on highways). Disagreements (e.g., box extents, class assignment) were resolved in technical review meetings, achieving consensus on the final labels. Annotations were exported in YOLO format (one TXT per image; normalized <class_id> <x_center> <y_center> <width> <height> with origin at the top-left corner; floating-point decimals, space-separated, no header). BR-AsPavDam (Dataset Summary & Organization) BR-AsPavDam comprises 2,167 images with 3,918 labeled defects, released for academic use. Final classes (after consolidation): Fissures (cracking) Shoving (merged: corrugation, shoving, rutting) Ravelling (loss of aggregate, surface ravel) Pothole Patch (visible maintenance/repair areas) Distribution of bounding boxes (high-resolution source set): Class # Boxes Fissures 1,843 Shoving 14 Ravelling 1,097 Pothole 125 Patch 839 Total 3,918 From the 3,565 initially extracted frames, 2,167 images remained after quality control (non-defective or low-quality frames removed). Folder structure (ready for object detection training): At the dataset root: images/ and labels/. images/ is organized by predominant normative class into five subfolders: Fissures/, Patch/, Pothole/, Ravelling/, Shoving/ (for Deformations). labels/ has no subfolders; for each image <name>.png, there is a homonymous <name>.txt file with YOLO boxes. File naming (predominant-class scheme):<PredominantClass>_NNN.png and <PredominantClass>_NNN.txt.If two or more classes occur with the same frequency in an image, the name follows a severity order (from most to least severe): Pothole > Patch > Shoving > Fissures > Ravelling. Naming reflects the predominant class only; all objects present are fully annotated in the corresponding TXT file. Image & label format: Images: PNG, 1152 × 2048 px (width × height) Labels: YOLO normalized format; one line per instance:<class_id> <x_center> <y_center> <width> <height> Class IDs (0–4):0 = Fissures, 1 = Shoving, 2 = Ravelling, 3 = Pothole, 4 = Patch Notes for reuse & reproducibility:— The dataset adheres to DNIT defect definitions and was curated by an expert labeling committee with consensus review.— Sampling choices (daylight window, ~2 s frame spacing, target 40 km/h, route diversity) were designed to balance visibility and non-redundancy.— Users should be aware of the class imbalance (e.g., few Deformations) typical of urban networks and consider appropriate training strategies (e.g., reweighting, augmentation).