A synthetically generated YOLO pose estimation dataset for arrow detection in medical images. This dataset consists of arrow annotations overlaid on medical images with pose keypoints marking the arrow tip and tail positions.

Dataset Overview

Total Images: Generated with 2x augmentation (each source image processed twice with different random arrows)
Train/Val/Test Split: 80/10/10
Format: YOLO Pose Estimation (v8 compatible)
Annotation Type: 2D Keypoint Pose with Bounding Boxes
Classes: 1 (Arrow)
Keypoints: 2 per object (tip, tail)

Dataset Structure

dataset/
├── data.yaml                 # YOLO configuration file
├── train/
│   ├── images/              # Training images (80% of data)
│   └── labels/              # Training labels in YOLO format
├── valid/
│   ├── images/              # Validation images (10% of data)
│   └── labels/              # Validation labels in YOLO format
└── test/
    ├── images/              # Test images (10% of data)
    └── labels/              # Test labels in YOLO format

Label Format

Each image has a corresponding .txt label file with the same name. Each line represents one arrow with the following format:

<class_id> <cx> <cy> <width> <height> <tip_x> <tip_y> <tip_vis> <tail_x> <tail_y> <tail_vis>

Where:

<class_id>: Object class (0 for arrow)
<cx>, <cy>: Normalized center coordinates of bounding box (0-1)
<width>, <height>: Normalized bounding box dimensions (0-1)
<tip_x>, <tip_y>: Normalized coordinates of arrow tip keypoint (0-1)
<tip_vis>: Tip visibility flag (0=not labeled, 1=labeled but not visible, 2=labeled and visible)
<tail_x>, <tail_y>: Normalized coordinates of arrow tail keypoint (0-1)
<tail_vis>: Tail visibility flag (0=not labeled, 1=labeled but not visible, 2=labeled and visible)

Note: All keypoints in this dataset have visibility=2 (labeled and visible) since they are synthetically generated.

Example label line:

0 0.512345 0.654321 0.250000 0.180000 0.645321 0.654321 2 0.378900 0.654321 2

Generation Details

Data Augmentation

Each source medical image is processed twice with:

Random arrow overlay positions (constrained to stay within image bounds)
Random arrow rotations (0-360°, with edge constraints)
Random arrow scaling (0.4x - 1.2x)
Random arrow colors (85% recolored, 10% original, 5% inverted)
Random number of arrows per image (Gaussian distribution: mean=3, std=2, range=[0,8])
Anti-overlap detection (arrows don't overlap with existing arrows)

Image Processing

All images resized to 1024px width while preserving aspect ratio
Images saved in original format (jpg/png/bmp/tiff)
Randomized file order before train/val/test split to ensure balanced augmentations across splits

Usage with YOLOv8

Training

from ultralytics import YOLO

# Load a pretrained model
model = YOLO('yolov8n-pose.pt')

# Train the model
results = model.train(
    data='path/to/dataset/data.yaml',
    epochs=100,
    imgsz=1024,
    batch=16,
    device=0
)

Validation

# Validate the model
metrics = model.val()

Inference

# Predict on an image
results = model.predict(source='image.jpg')

# Visualize results
for result in results:
    result.show()  # Display predictions

Dataset Configuration (data.yaml)

The data.yaml file contains:

path: dataset
train: train/images
val: valid/images
test: test/images
nc: 1
names:
  0: arrow
kpt_shape: [2, 3]  # 2 keypoints, 3 values each (x, y, visibility)

Keypoint Information

Keypoint 0 (Tip): The pointed end of the arrow
Keypoint 1 (Tail): The tail/base end of the arrow

Both keypoints are normalized to the image dimensions (0-1 range).

Statistics

Arrow Count Distribution

Mean arrows per image: ~3
Min arrows: 0
Max arrows: 8
Distribution: Gaussian (centered at 3)

Image Sizes

Width: Fixed at 1024px
Height: Variable (preserves aspect ratio of source images)
Format: JPG/PNG/BMP/TIFF

Citation

If you use this dataset, please cite:

@dataset{synthetic_medical_arrows_2026,
    title={Synthetic Medical Arrows Dataset},
    author={Kevin Xiao},
    year={2026},
    url={https://github.com/xckevin/synthetic-medical-arrows}
}

License

MIT

Notes

Images are resized to 1024px width for consistent training
The 2x augmentation strategy ensures diverse arrow placements and appearances
Randomization before splitting ensures train/val/test sets have balanced augmented samples
The dataset is generated synthetically, so real-world performance may vary

Troubleshooting

Images not found during training

Ensure paths in data.yaml are relative to the dataset root or absolute paths
Verify that image files exist in the train/images, valid/images, and test/images directories

Label format errors

Verify that label files have the same name as images (different extension)
Check that all coordinate values are normalized (between 0 and 1)
Ensure each line has exactly 11 space-separated values (class, bbox, keypoints with visibility)
Visibility flags should be 0, 1, or 2 (this dataset uses 2 for all keypoints)

Low model performance

Try increasing augmentation strength during training (brightness, contrast, etc.)
Consider increasing the number of source images or augmentation factor
Verify annotations are accurate by visualizing a few samples

Downloads last month: 9

Total file size:

685 MB