PopScore

Intrinsic Image Popularity Assesment

This is a browser-based implementation of the Intrinsic Image Popularity Assessment model that predicts how popular an image would be on social media based on its visual content and appeal.

Web App

PopScore.app

Features

• Drag & Drop Interface: Simply drag and drop images or browse files
• Real-time Processing: Uses the actual trained ResNet-50 model converted to ONNX format
• Visual Results: Shows popularity scores with intuitive color-coding and explanations
• Responsive Design: Works on desktop and mobile devices
• Privacy-First: All processing happens locally in your browser

How to Use

1. Upload an Image:

   • Drag and drop an image file onto the upload area, or
   • Click to browse for an image

   

2. Analyze:

   • Click "What's my PopScore" button to run the prediction

3. View Results:

   • See your image's popularity score and category
   • Read insights about what makes images popular

Technical Details

Model

• Architecture: ResNet-50 with modified final layer (2048 → 1 output)
• Training: Trained on 2.5M Instagram image pairs using learning-to-rank
• Accuracy: 76.65% pairwise accuracy, surpassing human-level performance
• Format: ONNX model for browser compatibility

Preprocessing

• Images are resized to 224×224 pixels
• Pixel values are normalized to [0, 1] range
• No additional normalization

Score Interpretation

• Excellent (4.0-6.0): Exceptional viral potential
• Good (2.0-4.0): Strong engagement potential
• Fair (0.0-2.0): Moderate appeal
• Poor (< 0.0): May struggle to gain traction

Browser Compatibility

• Chrome/Edge: Full support
• Firefox: Full support
• Safari: Full support
• Mobile: Responsive design works on all modern mobile browsers

Privacy & Security

• No Data Upload: All processing happens locally in your browser
• No Tracking: No analytics or user tracking
• Secure: Uses HTTPS-compatible ONNX.js runtime

Performance

• Model Loading: ~2-5 seconds on first visit
• Inference Time: ~1-3 seconds per image
• Memory Usage: ~200MB for model weights
• File Size Limit: 10MB per image

Research Context

Based on the paper "Intrinsic Image Popularity Assessment" by Ding et al., this implementation focuses on predicting intrinsic popularity (visual content only) rather than absolute popularity (which includes social factors like follower count, hashtags, etc.).

The model identifies visual patterns that correlate with engagement:

• People and faces
• High-quality photography
• Interesting compositions
• Scenic landscapes
• Cute animals

Files Structure

web/
├── wasm/              # Onix web WebAssembly files
├── index.html         # Main web interface
├── style.css          # Styling and responsive design
├── script.js          # JavaScript application logic
├── ort.min.js         # Onix file loader logic
├── model.onnx         # Converted ResNet-50 model
├── LICENCE            # MIT Licence
└── README.md          # This file

Troubleshooting

Model Won't Load

• Check browser console for errors
• Try refreshing the page

Slow Performance

• Model loading is normal on first visit
• Subsequent analyses are faster
• Consider using a faster device or browser

Inaccurate Results

• Model works best on clear, well-lit images
• Very dark or blurry images may not score well
• Results are predictions, not guarantees

Credits

• Original Research: "Intrinsic Image Popularity Assessment" by Keyan Ding, Kede Ma, and Shiqi Wang