AdithyaSK/cursor-detection-synthetic-dataset

Cursor Detection Synthetic Dataset

A synthetic object detection dataset for mouse cursor detection, generated by compositing real cursor images onto website screenshots.

Dataset Structure

Dataset({
    features: ['image', 'bbox', 'split'],
    num_rows: 650
})

Feature	Type	Description
image	PIL Image	640x640 RGB image with cursor composited
bbox	string (JSON)	YOLO-format bounding box: {class, x_center, y_center, width, height}
split	string	train / val / test

Splits

Split	Images
train	500
val	100
test	50

Dataset Generation

The dataset was generated by:

1. Cursors: 366 cursor types from Fraser/cursors — includes various OS themes (macOS, Windows, Linux) and cursor states (default, pointer, text, grab, etc.) with alpha channels and hotspot metadata.

2. Backgrounds: 1,688 website screenshots from naorm/website-screenshots.

3. Augmentation: Each image was created by:

   • Randomly selecting a screenshot background
   • Randomly selecting a cursor type
   • Resizing cursor to 16–48px
   • Placing it at a random position using alpha compositing with correct hotspot alignment
   • Generating a normalized YOLO bounding box label

Usage

from datasets import load_dataset
import json

ds = load_dataset("AdithyaSK/cursor-detection-synthetic-dataset", split="train")

sample = ds[0]
img = sample["image"]
bbox = json.loads(sample["bbox"])
print(f"Cursor at: ({bbox['x_center']}, {bbox['y_center']}) size {bbox['width']}x{bbox['height']}")

Convert to YOLO format

import os
from datasets import load_dataset

def save_yolo_format(output_dir="cursor_yolo"):
    ds = load_dataset("AdithyaSK/cursor-detection-synthetic-dataset")
    
    for split in ["train", "val", "test"]:
        img_dir = os.path.join(output_dir, "images", split)
        lbl_dir = os.path.join(output_dir, "labels", split)
        os.makedirs(img_dir, exist_ok=True)
        os.makedirs(lbl_dir, exist_ok=True)
        
        for i, sample in enumerate(ds[split]):
            # Save image
            sample["image"].save(os.path.join(img_dir, f"{i:06d}.jpg"))
            # Save label
            bbox = json.loads(sample["bbox"])
            with open(os.path.join(lbl_dir, f"{i:06d}.txt"), "w") as f:
                f.write(f"{bbox['class']} {bbox['x_center']} {bbox['y_center']} {bbox['width']} {bbox['height']}\n")

save_yolo_format()

License

Cursor images from Fraser/cursors are under their respective licenses. Website screenshots are from naorm/website-screenshots.