A while ago I played with style visualizations and bilateral filters. The latter have the nice property of filtering out noise but preserving edges. Here are some example class from GoogLeNet (Inception network). Big shout out to Audun m. Øygard and Kyle McDonald who were among the first to use filters (e.g. gaussian blurs) essentially as image regularizers for single class visualizations. These visualizations here were directly inspired by their ideas.
Neural Network art show
I’ll be showing 3 high-res pieces based on this technique at the Neural Network art show at Gray Area in San Francisco end of February.
Class #34, Leatherback Turtle
Class #76, Tarantula
Class #144, Pelican
Class #156, Blenheim Spaniel
Class #488, Chain
Class #725, Pitcher
Class #944, Artichoke
Fun with video
Here’s a little animation I call the “Animal parade”
Combining bilateral filters with style transfer
These were made using bilateral class visualization using GoogLeNet (cello, violin, saxophone) and then had style transfer applied with the style coming from a natural picture (photo).