farazv2/new-overlay-model-yolo

YOLO Overlay Detection Model - Optimized

This model was trained to detect and segment overlay elements in images/videos using YOLOv8 segmentation with optimized hyperparameters.

Model Details

• Model Type: YOLOv8 Instance Segmentation
• Architecture: auto
• Framework: Ultralytics YOLO
• Training Date: 2025-11-07
• Task: Instance Segmentation
• Classes: Overlay elements
• Image Size: 800px (optimized for detail)

Performance Metrics

Metric	Value
Box [email protected]	0.9038
Box [email protected]:0.95	0.7171
Mask [email protected]	0.3981
Mask [email protected]:0.95	0.1520

Key Optimizations

This model includes several optimizations over the baseline:

• ✅ Mosaic Augmentation enabled (1.0) - Critical for YOLO performance
• ✅ Copy-Paste Augmentation (0.3) - Essential for segmentation tasks
• ✅ Larger Image Size (800px) - Better detail capture
• ✅ Cosine LR Scheduler - Smoother convergence
• ✅ Multi-Scale Training - Better scale invariance
• ✅ Enhanced Augmentations - Rotation (10°), Scale (0.5), Perspective
• ✅ Optimized Batch Size (32) - Better gradient estimates on dual GPUs

Usage

Installation

pip install ultralytics

Inference

from ultralytics import YOLO
from huggingface_hub import hf_hub_download

# Download model
model_path = hf_hub_download(
    repo_id="farazv2/overlay-model-yolo",
    filename="best.pt"
)

# Load model
model = YOLO(model_path)

# Run inference
results = model('image.jpg')

# Process results
for result in results:
    boxes = result.boxes  # Bounding boxes
    masks = result.masks  # Segmentation masks
    
    # Visualize
    result.show()
    
    # Save
    result.save('output.jpg')

Batch Inference

# Process multiple images
results = model(['image1.jpg', 'image2.jpg', 'image3.jpg'])

# Process video
results = model('video.mp4', save=True)

Training Configuration

Parameter	Value	Notes
Epochs	10
Image Size	800	Increased from 640
Batch Size	16	Optimized for dual T4
Optimizer	AdamW
Initial LR	0.0005	With cosine scheduler
Mosaic	1.0	Re-enabled (critical!)
Copy-Paste	0.3	New addition
Multi-Scale	True	Enabled
Mixed Precision	True	Enabled
Patience	25

Model Export

The model can be exported to various formats:

from ultralytics import YOLO

model = YOLO('best.pt')

# Export to ONNX
model.export(format='onnx')

# Export to TensorRT
model.export(format='engine')

# Export to CoreML
model.export(format='coreml')

Citation

If you use this model, please cite:

@software{overlay_yolo_model,
  author = {farazv2},
  title = {YOLO Overlay Detection Model - Optimized},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/farazv2/overlay-model-yolo}
}

License

This model is released under the AGPL-3.0 license, following Ultralytics YOLOv8 licensing.

Acknowledgments

• Built with Ultralytics YOLOv8
• Trained on Kaggle with GPU acceleration
• Optimized with best practices for segmentation tasks