ArnavvvvK/rfdetr

RF-DETR Face Detection Models

Optimized RF-DETR (Receptive Field DETR) models for face detection with pruning and quantization techniques.

🚀 Quick Start

Using the Models

import onnxruntime as ort
from PIL import Image
import torchvision.transforms as transforms

# Load model
session = ort.InferenceSession("rf_detr_quantized.onnx")

# Preprocess image
transform = transforms.Compose([
    transforms.Resize((800, 800)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

image = Image.open("image.jpg").convert("RGB")
input_tensor = transform(image).unsqueeze(0).numpy()

# Run inference
input_name = session.get_inputs()[0].name
outputs = session.run(None, {input_name: input_tensor})

# Process results
logits = outputs[0][0]
boxes = outputs[1][0]
max_logits = logits.max(axis=-1)

# Extract detections
detections = []
for i in range(len(boxes)):
    if max_logits[i] > 0.3:  # Confidence threshold
        box = boxes[i] * 800  # Scale to image size
        detections.append({
            'bbox': box.tolist(),
            'confidence': float(max_logits[i])
        })

print(f"Found {len(detections)} faces")

Command Line Demo

python inference_demo.py --model rf_detr_quantized.onnx --image test.jpg --output result.jpg

📊 Model Performance

🎯 Key Features

• Multiple Variants: Original, pruned (1%, 30%), and quantized models
• Optimized for Deployment: ONNX format for fast inference
• Size Reduction: Up to 74% smaller with quantization
• High Accuracy: F1 score of 0.990 on WIDER FACE dataset
• Easy Integration: Simple Python API

📁 Available Models

• rf_detr.onnx - Baseline model (19MB)
• rf_detr_pruned_1.onnx - 1% pruned model (19MB)
• rf_detr_pruned_30.onnx - 30% pruned model (19MB)
• rf_detr_quantized.onnx - Quantized model (5MB) ⭐ Recommended

🔧 Requirements

pip install onnxruntime torchvision Pillow numpy

📈 Benchmark Results

Performance vs Model Size Trade-off

F1 Score vs Model Size

🖼️ Example Detections

🏗️ Model Architecture

• Backbone: RF-DETR (Receptive Field DETR)
• Input: 800×800 RGB images
• Output: Bounding boxes + confidence scores
• Optimization: L1 pruning + 8-bit quantization

📚 Citation

@article{rfdetr2023,
  title={RF-DETR: Receptive Field DETR for Object Detection},
  author={...},
  journal={...},
  year={2023}
}

📄 License

MIT License - see LICENSE file for details.

🤝 Contributing

Feel free to open issues or submit pull requests for improvements.

Recommended Usage: Use rf_detr_quantized.onnx for deployment - it's 74% smaller with the same performance!