---
license: agpl-3.0
language:
- en
pretty_name: HSRP/NON-HSRP Dataset
task_categories:
- image-classification
tags:
- HSRP
- traffic
- license_plate
- indian_vehicles
- computer-vision
size_categories:
- 1K<n<10K
---
# HSRP Classification Dataset

## Description

Dataset created for the Smart-HSRP project to train an image
classification model for HSRP-related classification.

## Ownership

This dataset was collected, prepared, and annotated for this project.

## Classes

| HSRP | 0 |
|-------|------|
| NON-HSRP | 1 |

## Dataset Structure
```text
train/
├──hsrp/
└──non-hsrp/

validation/
├──hsrp/
└──non-hsrp/

test/
├──hsrp/
└──non-hsrp/
```
## Collection

All the images were either manually collected or stock images.

## Annotation

All the images were manually annotated.

## Preprocessing

The dataset consists of license-plate image crops extracted from larger source images using a YOLO-based plate detection model.

For each source image, the image is divided into non-overlapping 640 × 640 pixel tiles. A stride of 640 pixels is used, so there is no overlap between adjacent tiles. Tiles with a height or width smaller than 200 pixels are skipped.

The plate detector (`plate_detector.pt`) is applied to each tile with a confidence threshold of 0.20. Detected bounding boxes are converted from tile coordinates back to the coordinates of the original full-resolution image.

The detected plate regions are then cropped directly from the original image. Bounding-box coordinates are clipped to the image boundaries before cropping. Empty crops are discarded, and valid crops are saved as JPEG (`.jpg`) images.

The resulting cropped license-plate images form the input dataset for the subsequent HSRP classification stage.

No additional resizing or normalization is performed by this cropping pipeline. Any resizing, normalization, or other preprocessing applied specifically during HSRP classifier training is not documented here unless explicitly specified by the classifier training configuration.

## Splits

The dataset is organized into training, validation, and test splits.

Exact sample counts for each split should be reported here:

| Split | HSRP | NON-HSRP |
|---|---|---|
| Train | 1,506x | 1,405x |
| Validation | 225x | 210x |
| Test | 227x | 212x |
| **Total** | **1,958** | **1,827** |

The split counts should correspond to the final dataset version uploaded to this repository.

## Intended Use

This dataset is intended for training and evaluating image classification models for HSRP-related license-plate classification as part of the Smart-HSRP project.

The dataset is designed to be used after a license-plate detection stage, where license plates are detected in larger images and cropped from their original image coordinates.

Potential applications include:

- HSRP-related license-plate classification
- Automated vehicle and license-plate analysis
- Traffic monitoring systems
- Computer vision research
- Smart-HSRP detection pipelines

The dataset is intended for research and development and should be evaluated for the specific deployment environment before being used in production or safety-critical applications.

## Limitations

The dataset contains cropped license-plate images generated using a YOLO-based plate detection pipeline. Consequently, the quality and characteristics of the dataset depend partly on the performance of the plate detector used during data preparation.

Potential limitations include:

- Incorrect or missed plate detections can result in missing samples or unsuitable crops.
- Bounding-box errors from the plate detector may cause portions of the plate to be cropped incorrectly.
- Image quality may vary depending on the original source images.
- Performance may be affected by differences in lighting, camera angle, image resolution, motion blur, occlusion, and viewing distance.
- The dataset may not represent all HSRP formats, vehicle types, geographic regions, camera configurations, or real-world conditions.
- Models trained on this dataset may experience reduced performance when applied to images that differ substantially from the data distribution used during dataset creation.
- The dataset should not be assumed to provide complete coverage of all possible HSRP appearances or real-world traffic conditions.

The dataset should therefore be independently evaluated on representative data before deployment in a production system.

## Related Model

Smart-HSRP HSRP Classifier:[![HSRP classifer model](https://img.shields.io/badge/Huggingface-HSRP%20Efficientnet%20B0-yellow?logo=huggingface)](https://huggingface.co/Smolry/HSRP-classification)

## Related Project

Smart-HSRP:[![Smart-HSRP-detection](https://img.shields.io/badge/Github-Smart%20HSRP%20Detection-blue?logo=github)](https://github.com/Smolry/Smart-HSRP-detection)