BogieSeg: A Dataset for Railway Bogie Component and Wheel Defect Detection

BogieSeg is a high-resolution annotated image dataset designed for deep learning-based automated detection and instance segmentation of railway bogie components and external wheel defects. It is the first publicly available benchmark dataset that simultaneously covers both bogie sub-assembly components and external wheel surface defects from a real-world operational railway environment.

The dataset was collected at Kotri Railway Junction, Pakistan, using an indigenously designed camera stand equipped with a GoPro Hero 9 (3840 × 2160 pixels, 30 FPS). The system was deployed trackside to capture moving trains at operational speeds of 10–40 km/h, providing realistic and challenging image conditions including variable lighting, motion blur, and surface corrosion.

Dataset Composition

Total images: 905 high-resolution images (augmented from 391 manually annotated frames)

Source video frames extracted: ~8,000 (from multiple train passages)

Frames selected for annotation: 391

Image resolution: 3840 × 2160 pixels

Annotation type: Instance segmentation (polygon annotations) + bounding boxes

Classes (9 categories)

Class
Description

Axle Box
Bearing housing connecting axle to bogie frame

Binding Screw
Fastening component in bogie assembly

Connecting Rod
Structural linkage between bogie parts

Crack / Scratch
Surface defects on wheel or metal components

Nut
Small fastening component

Primary Suspension Spring
Main vibration-absorbing spring element

Support Beam
Main structural support of the bogie

Support Rod
Secondary structural support element

Wheel
Railway wheel including tread surface

Dataset Structure

The dataset is provided in the following forms:

Original images — Raw high-resolution frames before augmentation

Labeled / Annotated images — Polygon-annotated instance segmentation images

Annotation Formats

Ground truth annotations are available in five standard formats:

TXT — YOLO format (normalized bounding box / polygon coordinates + class label)

XML — PASCAL VOC format

JSON — COCO format (instance segmentation with polygon coordinates)

CSV — Tabular format with class information and coordinates

TFRecord — TensorFlow native binary format

This makes BogieSeg natively compatible with major DL frameworks including YOLOv8, Detectron2, Mask R-CNN, FastInst, SparseInst, and vision transformer-based models.

Data Split

Split
Images
Proportion

Training
634
70%

Validation
181
20%

Test
90
10%

Stratified sampling was used to ensure proportional class representation across all three splits.

Data Augmentation

To improve model generalization, 391 manually annotated images were augmented to 905 images using the following transformations (applied to both images and annotations):

Cropping (horizontal and vertical)

Horizontal and vertical flipping

Brightness adjustment

Shear (horizontal and vertical)

Rotation (90°, 180°, 270°)

Hue shifting

Saturation variation (±10%)

Exposure adjustment

Blur

Gaussian noise

Baseline Model Performance

The dataset was validated by training YOLOv8-seg (instance segmentation variant):

Metric
Result

Overall Accuracy
93%

mAP@0.5
88%

Axle Box / Binding Screw / Connecting Rod / Support Beam / Wheel
100% each

Primary Suspension Spring
98%

Nut
94%

Crack / Scratch
78%

Training hardware: Intel Core i7 (9th Gen), 16 GB RAM, NVIDIA GeForce RTX 2060 Super (8 GB VRAM).

Data Collection Details

Location: Kotri Railway Junction, Sindh, Pakistan

Camera: GoPro Hero 9, mounted 3 inches from ground and 10 inches from track centerline

Speed of trains: 10–40 km/h

Trains covered: 10 different trains (5 Up / 5 Down), including Karachi Express, Karakoram Express, Mehran Express, Millat Express, and Pak Business Express

Collection period: Afternoon to evening (post 12:00 hrs) to capture variable natural and artificial lighting

Government authorization: Collected with permission from Pakistan Railways and the Government of Pakistan

Annotations were performed using Roboflow and validated in direct collaboration with railway inspection personnel from the Pakistan Railways department.

Relation to Prior Work

BogieSeg extends the authors' previously published FaultSeg dataset (Scientific Data, 2025), which focused solely on wheel defect detection. BogieSeg adds complete bogie sub-assembly components alongside wheel and surface defects, enabling a comprehensive rolling stock inspection benchmark.

FaultSeg on Zenodo: https://doi.org/10.5281/ZENODO.13162335

FaultSeg on Figshare: https://springernature.figshare.com/articles/dataset/FaultSeg_A_Dataset_for_Train_Wheel_Defect_Detection/27996866

Applications

Training and benchmarking object detection and instance segmentation models for railway inspection

Automated bogie component identification and defect localization

Predictive maintenance systems for rolling stock

Safety protocol development for railway operations

Transfer learning for related industrial inspection tasks

License

This dataset is released under an open license for educational and research purposes.

Citation

If you use this dataset, please cite the associated paper:

Muhammad Zakir Shaikh, Enrique Nava Baro, Bushra Abro, Elidia Beatriz Blázquez-Parra, Sahil Jatoi, and Bhawani Shankar Chowdhry. BogieSeg: A Dataset for Railway Bogie Component Detection and Wheel Defect Instance Segmentation. [Journal name], [Year].

Authors and Affiliations

Muhammad Zakir Shaikh — National Center for Robotics, Automation and Artificial Intelligence (NCRAAI), Mehran University of Engineering and Technology (MUET), Pakistan; University of Malaga, Spain

Enrique Nava Baro — Departamento de Ingeniería de Comunicaciones, Universidad de Malaga, Spain

Bushra Abro — NCRAAI, MUET, Pakistan

Elidia Beatriz Blázquez-Parra — Department of Graphical Engineering, Design and Projects, Universidad de Malaga, Spain

Sahil Jatoi — NCRAAI, MUET, Pakistan

Bhawani Shankar Chowdhry — NCRAAI, MUET, Pakistan

Corresponding author: [e-mail]

Acknowledgements

This research was supported by the Departamento de Ingeniería de Comunicaciones, Universidad de Malaga (Spain); National Center for Robotics, Automation and Artificial Intelligence (NCRAAI MUET); Higher Education Commission Pakistan; Sindh Higher Education Commission Pakistan; and the EU Funded Erasmus Plus Capacity Building in Higher Education ACTIVE Climate Action Project and CENTRAL Project (ID: 598914). Special thanks to Pakistan Railways, Kotri Railway Station, and the Carriage and Wagon Workshop, Hyderabad for their support during data collection.