RoadAb-India 2026: A Multi-Layer Road Abnormality Dataset with Object Annotations, Severity Labels, and Geospatial Context

RoadAb-India 2026 is a multi-layer road abnormality image dataset developed to support computer vision and metadata-aware research on road-condition assessment. The dataset contains 2,074 road-scene images collected from multiple geographic regions in India, covering Tamil Nadu, Delhi, and Maharashtra. The images represent urban, residential, rural, and highway road environments under varied lighting, weather, traffic, pavement, and seasonal conditions.

The dataset includes 3,391 object-level annotations distributed across 1,984 images containing visible road abnormalities. An additional 90 images contain no visible road abnormalities and are intentionally retained as negative road-scene samples. These negative samples can support false-positive analysis, background modelling, robustness assessment, and realistic evaluation of road-abnormality detection systems.

Visible road abnormalities are represented using normalized object-detection bounding boxes. The dataset covers six abnormality classes:

1. Surface depression
2. Pothole
3. Manhole
4. Road patch failure
5. Crack
6. Miscellaneous abnormality

The dataset is organized into four complementary metadata layers:

Layer 0 — Raw Image Metadata

Contains one record for each of the 2,074 images, including the image identifier, file format, file size, image resolution, collection date, and collection state.

Layer 1 — Object Annotation Metadata

Contains 3,391 object-level annotation records. Each record identifies an annotated road abnormality and provides its image identifier, abnormality identifier, abnormality class, normalized bounding-box centre coordinates, normalized width, and normalized height.

Layer 2 — Semantic Metadata

Contains one record for each image and provides semantic and environmental attributes, including abnormality type, severity, road type, weather condition, lighting condition, occlusion level, and abnormality shape. Severity is categorized as minor, moderate, or severe. Abnormality-specific fields are intentionally empty for the 90 negative road-scene images.

Layer 3 — Geospatial and Road-Context Metadata

Contains one record for each image and provides collection state, district, latitude and longitude grid values, pavement type, traffic density, collection season, nearby road features, slope type, road-width category, and proximity to an intersection.

The dataset supports research tasks including:

• Road abnormality object detection
• Road abnormality classification
• Multi-label road-scene classification
• Abnormality severity prediction
• Metadata-aware computer vision
• Context-aware road-condition assessment
• Geographic and environmental robustness analysis
• Dataset bias and domain-shift analysis
• Negative-sample and false-positive evaluation
• Multi-task and multimodal machine learning

The images were manually reviewed and annotated using a consistent road-abnormality labelling protocol. Bounding boxes were designed to cover the visible extent of each abnormality. A quality-control review was conducted on a randomly selected subset of 207 images. Inter-annotator evaluation on the reviewed subset produced a mean bounding-box Intersection over Union of 0.91, class agreement of 96.80%, and Cohen’s kappa of 0.94.

All identified annotation corrections were incorporated before the preparation of this release. Duplicate and near-duplicate captures were reviewed and retained only where they represented naturally occurring repeated observations or different viewpoints. Related images were assigned to the same dataset partition during evaluation to avoid train–validation–test leakage.

RoadAb-India 2026 is intended to complement larger road-damage datasets by emphasizing annotation richness, severity information, environmental context, and fine-grained geospatial metadata rather than dataset scale alone.

The dataset is initially distributed under restricted access while the associated data descriptor manuscript is under preparation and peer review. Access conditions may be updated after publication of the associated article.

Dataset Contents

The deposited files include:

• roadab_india_2026_images.zip — 2,074 road-scene images.
• roadab_india_2026_annotations.zip — 2,074 image-level annotation files, including empty annotation files for negative road scenes.
• layer_0_raw_image_metadata.csv — Image-level technical and collection metadata.
• layer_1_annotation_metadata.csv — Object-level abnormality labels and normalized bounding boxes.
• layer_2_semantic_metadata.csv — Severity, environmental, road-context, occlusion, and shape attributes.
• layer_3_geospatial_metadata.csv — Geographic, pavement, traffic, seasonal, and road-configuration attributes.
• README.md — Dataset organization, class definitions, file relationships, annotation format, and usage instructions.

File and Record Relationships

The image_id field is the primary identifier connecting the images and all four metadata layers.

Layer 0, Layer 2, and Layer 3 contain one record for every image. Layer 1 contains one row for every annotated abnormality object and can therefore contain multiple rows for the same image.

Images containing no visible road abnormality have no object rows in Layer 1. Their corresponding annotation files should be retained as empty files to preserve one-to-one alignment between image and annotation filenames.

Annotation Format

Object annotations use normalized bounding-box coordinates:

• x_centre: horizontal centre of the bounding box, normalized by image width.
• y_centre: vertical centre of the bounding box, normalized by image height.
• width_scaled: bounding-box width normalized by image width.
• height_scaled: bounding-box height normalized by image height.

All normalized coordinate values are intended to fall between 0 and 1.

Geographic Coverage

India:

• Tamil Nadu
• Delhi
• Maharashtra

District-level metadata includes Chennai, Chengalpattu, Tiruvallur, North West Delhi, and Sangli.