Skip to main content

Style Transfer Progress

I made a longer version of the prototype I showed in my Midterm Status post. This took a week to complete:

Overall I'm happy with this. I really like how the frames of the animation are coherent. There's no flickering from one frame to the next. I like how the style augments the scene.

I don't like the animation's stability. Some of the frames jump around a bit. It isn't that bad here but it will be worse with some of the other Google Street View scenes. I've been working on improving this in my code. I have made a lot of progress but am not completely satisfied with my solution. I have an idea that I believe will work for both planar and equirectangular frames. I will finish coding it tomorrow during class.

The animation took a week to create with resolution of only 1000x500. The final version will need to be much higher than that. To start I used Neural Style Transfer code I found online but am eager to code my own. This code uses an algorithm outlined in the seminal paper Image Style Transfer Using Convolutional Neural Networks by Gatys, Ecker, and Bethge. I've been reading other research papers and found some that claim to be a lot faster. I am going to experiment with them in the short term. When I find one that I like I will re-implement it myself and adapt it for my purposes.


Comments powered by Disqus