# Cursor Detection YOLOv8n

A YOLOv8n model trained to detect mouse cursors in screenshots and video frames.

## Training Details

- **Base Model:** YOLOv8n (3.2M parameters)
- **Training Data:** Synthetic dataset generated by compositing 366 different cursor types onto 1688 website screenshots
- **Dataset Size:** 500 train / 100 val / 50 test
- **Image Size:** 640x640
- **Epochs:** 30
- **Hardware:** NVIDIA T4 GPU

## Performance

| Metric | Value |
|--------|-------|
| mAP50 | 92.1% |
| mAP50-95 | 58.2% |
| Precision | 84.8% |
| Recall | 89.5% |

## Dataset Generation

The synthetic dataset was created by:
1. Loading cursor images from [Fraser/cursors](https://huggingface.co/datasets/Fraser/cursors) (366 cursor types with hotspot info)
2. Loading background screenshots from [naorm/website-screenshots](https://huggingface.co/datasets/naorm/website-screenshots)
3. Compositing cursors at random positions with alpha blending
4. Generating YOLO format bounding box labels

## Usage

```python
from ultralytics import YOLO

# Load model
model = YOLO("AdithyaSK/cursor-detection-yolov8n/best.pt")

# Detect cursor in an image
results = model("screenshot.jpg")
results[0].show()
```

## License

AGPL-3.0 (same as Ultralytics YOLOv8)