rf-detr-stamp-signature-detector

RF-DETR Stamp/Signature/Checkbox Detector

A fine-tuned RF-DETR model for detecting stamps, signatures, and checkboxes in document images.

Model Details

• Base Model: RF-DETR Base (DINOv2 backbone)
• Training: 30 epochs on 8x NVIDIA H100 80GB
• Dataset: ~9,400 images (merged from multiple sources)
• License: Apache 2.0

Classes

Performance

Usage

from rfdetr import RFDETRBase
from PIL import Image

# Load model
model = RFDETRBase()
model.load("path/to/checkpoint_best_ema.pth")

# Run inference
image = Image.open("document.png")
detections = model.predict(image, threshold=0.5)

for det in detections:
    print(f"Class: {det['class']}, Confidence: {det['confidence']:.2f}, Box: {det['bbox']}")

Training Details

• Batch size: 32 per GPU (256 effective)
• Learning rate: 4e-4
• Optimizer: AdamW
• Resolution: 560px

Limitations

• checkbox_unchecked detection is weaker (65.7% mAP) due to low visual distinctiveness
• Best suited for stamps and signatures in document images

Citation

@misc{bluecopa-stamp-detector,
  title={RF-DETR Stamp/Signature Detector},
  author={BlueCopa},
  year={2024},
  publisher={HuggingFace}
}