PakVehicle-Net: Car Make Classification Dataset (Pakistan Roads)

Overview

This dataset was developed to support car make classification on vehicles commonly found on Pakistani roads. It contains ~15,000 images across 49 classes (post-augmentation), annotated and prepared through Roboflow. The dataset was used to train a YOLOv11x-cls classification pipeline, combining object detection and classification for real-time video analysis.

---

Dataset Structure

• Total images: ~15,000 (post-augmentation)
• Number of classes: 49 (48 car makes + 1 Unknown class)
• Format: YOLO / COCO-style folder structure (train/, valid/, test/)
• Annotation tool: Roboflow UI

Splits (Pre-augmentation)

Split	Images	Notes
Train	~70%	Used for model training
Valid	~16%	Used for validation during training
Test	~14%	Held out for final evaluation

Splits (Post-augmentation)

Split	Images	Notes
Train	~87.5%	Used for model training
Valid	~6.7%	Used for validation during training
Test	~5.8%	Held out for final evaluation

The post-augmentation splits were actually used in the dataset validation training pipeline

---

Classes

• 48 specific car makes (e.g., Toyota Corolla, Suzuki Alto, Honda Civic, etc.)
• Unknown: 276 images (pre-augmentation) of random cars not included in the 48 classes, introduced to improve robustness against out-of-distribution samples.

---

Annotation Process

1. Initial Dataset:

• Sourced ~4,000 images from Roboflow (48 classes).
• Cleaned mislabeled and low-quality samples.

2. Expansion:

• Added ~3,000 additional images from PakWheels, OLX, and Google Images.
• Annotated manually using Roboflow UI.

3. Augmentation: Applied within Roboflow on training set only to improve model generalization:

• Random flips and rotations
• Brightness and contrast adjustment
• Blur / noise injection
• Scaling and cropping

4. Final Dataset Size: ~15,000 images across 49 classes.

---

Preprocessing & Pipeline Usage

The dataset was primarily used with the following pipeline:

1. Detection: Pretrained YOLOv11x detection model (cars only).
2. Cropping: Cars cropped using bounding box coordinates.
3. Classification: Cropped cars passed into YOLOv11x-cls classifier.
4. Annotation: Results re-attached to video frames with class + confidence overlay.

---

Known Issues / Notes

• Without the Unknown class, the classifier biased heavily towards Suzuki Carry for unrecognizable cars.
• Performance drops on low-resolution or angled images; further dataset expansion may improve this.

---

Citation

If you use this dataset in your research or project, please cite:

@dataset{pakvehicle-net, title = {PakVehicle-Net: Pakistani Car Make Classification Dataset}, author = {Muhammad Abdullah, Zia ur Rehman, Sharjeel Farooq, Abdul Moiz, Shaheer Murtaza Shehri, Abdullah Farooq}, year = {2025}, howpublished = {Hugging Face Datasets}, url = { https://huggingface.co/datasets/em4bm4lik/PakVehicle-Net } }

---

License

• This dataset is for educational and research purposes only.
• Images were collected from publicly available sources (Roboflow, PakWheels, OLX, Google Images).
• Redistribution may be subject to the original platforms’ terms of service.

---

Acknowledgements

• Roboflow for providing the annotation platform and initial dataset.
• Open-source contributors of YOLOv11.
• Our project supervisor for guidance.