Reviewing production OpenEXR files on the web for ML

Production images?

• OpenEXR files
• High resolution, 1080p and up
• Extended / High Dynamic Range
• 16-bit or 32-bit floating point
• Not only final composed color data
  • Alpha masks
  • Feature buffers from the renderer
  • Error maps
  • ...

Example Use Case: Denoiser

• Kernel-Predicting Convolutional Networks for Denoising Monte Carlo Renderings, Rousselle et al, ACM SIGGRAPH 2017
• Deep learning for denoising rendered images
• Suitable for production

Introducting: JERI.io

• The JavaScript Extended-Range Image Viewer
• HTTP server directly serves OpenEXR images with config
• Pure client-side app to decode EXR directly in the browser
• Thus, client receives original, unmodified pixels
• Like a very fancy, configurable, <img> tag for EXRs

EXR on the web

• OpenEXR compiled to WebAssembly
• Emscripten did the job
• Speed is fine for our use cases, in particular when caching
• Now we have EXRs in the browser!

Usable EXR on the web

Configuration

const config = {
  title: 'root',
  children: [
    { title: 'Improved', image: 'glass.exr' },
    { title: 'Reference', image: 'glass-ref.exr' },
    {
      title: 'SSIM',
      lossMap: {
        'function': 'SSIM',
        imageA: 'glass.exr',
        imageB: 'glass-ref.exr'
      }
    },
  ]
};

JERI Demo

Integration

https://jeri.io

Released as Open Source

Thank you