🔐 Revpass

YOLOv26s Fine-tuned for Google reCAPTCHA Detection

High-performance object detection model specialized for reCAPTCHA v2/v3 image recognition

---

📊 Performance Metrics

Metric	Score
mAP50	95.4%
mAP50-95	71.2%
Precision	89.7%
Recall	91.5%

Trained on 10,390 reCAPTCHA images (103 epochs, 1.46 hours, NVIDIA RTX PRO 6000 96GB)

---

🎯 Demo Results

Sample predictions on validation set showing high-confidence detections across all 11 classes

---

🏷️ Supported Classes (11)

bicycle, bridge, bus, car, chimney, crosswalk, 
fire_hydrant, motorcycle, palm_tree, stairs, traffic_light

---

🚀 Quick Start

Installation

pip install ultralytics

Usage

from ultralytics import YOLO

# Load model from HuggingFace
model = YOLO("hf://saifyxpro/Revpass")

# Run inference
results = model("captcha_tile.jpg", conf=0.25)

# Print predictions
for r in results:
    for box in r.boxes:
        class_name = model.names[int(box.cls[0])]
        confidence = float(box.conf[0])
        print(f"{class_name}: {confidence:.2%}")

Example Output

palm_tree: 99%
bus: 98%
traffic_light: 99%
fire_hydrant: 98%

---

🔧 Training Details

Parameter	Value
Base Model	YOLOv26s
Dataset	Google reCAPTCHA (Kaggle)
Training Images	8,832
Validation Images	1,558
Epochs	150 (stopped at 103)
Batch Size	64
Image Size	640x640
Optimizer	AdamW
Learning Rate	0.001 → 0.00001
GPU	NVIDIA RTX PRO 6000 (96GB)
Training Time	1.46 hours

Augmentation Strategy

• HSV color jittering
• Random translation & scaling
• Horizontal flipping
• Mosaic augmentation
• MixUp (10%)
• Copy-Paste (10%)

---

💻 CPU Optimization

This model is optimized for CPU inference while being trained on a high-end GPU for maximum quality:

• ✅ YOLOv26s architecture (small, fast)
• ✅ ONNX export support
• ✅ Efficient inference on consumer hardware
• ✅ No GPU required for deployment

Export to ONNX

model = YOLO("hf://saifyxpro/Revpass")
model.export(format="onnx", imgsz=640, simplify=True)

---

📈 Training Curves

The model achieved convergence at epoch 103 with early stopping (patience=30):

• Best mAP50: 95.4% (epoch 103)
• Final Loss: 0.42
• Validation Stability: High consistency in final 20 epochs

---

🎓 Use Cases

⚠️ EDUCATIONAL PURPOSE ONLY

This model is designed for research and educational purposes to demonstrate:

• Fine-tuning YOLO models on custom datasets
• Object detection for specialized domains
• High-performance training on large GPUs

Do not use for unauthorized access or bypassing security measures.

---

📦 Model Files

• best.pt - PyTorch weights (22.5 MB)
• data.yaml - Dataset configuration
• README.md - This file
• test_predictions.png - Demo results

---

🙏 Acknowledgments

• Ultralytics for the amazing YOLO framework
• Kaggle for hosting the reCAPTCHA dataset
• HuggingFace for model hosting infrastructure

---

📄 License

This project is licensed under AGPL-3.0. See LICENSE for details.

---

🔗 Links

• Model: HuggingFace Hub
• Framework: Ultralytics YOLO
• Dataset: Google reCAPTCHA (Kaggle)

---

Built with ❤️ using Ultralytics YOLOv26s