Monday, December 11, 2017

Image Translation

A new machine-learning algorithm can take a photo of a street scene and translate the image to another time of day or another weather condition. 

For example, the photo on the left shows is taken from a car on a rainy day. On the right, the computer translates the scene into a sunny day with a blue sky. 

Here the algorithm does the opposite, translating a photo of a sunny day (left) into a virtual image of the same scene in rainy conditions (right).

The night-to-day translations are impressive because there seems so little information to start with in the photo at left, and the change is so radical.

The system can also translate a photographic street scene into a graphic that looks like it comes from a video game — or it can take a still from a video game and make it look more photographic. 

It can also change the hair color of a person, or alter a dog from one breed to another. 

This machine-learning technology, driven by generative adversarial networks, is progressing very quickly, so any weaknesses or limitations we see in the results now will be overcome rapidly.

We can no longer say "Photos don't lie." 
Read More:
Google photo collection with lots more pairs of examples.
Scientists' paper as a PDF
Video Game Graphics To Reality And Back

Related Posts
Text-to-Image Synthesis
Generative design resembles Art Nouveau
Morphing Celebrities


robertsloan2art said...

That is pretty cool! I've done things like that from imagination and memory, but a program that does it accurately is awesome. Really useful for getting more information out of a night photo, I can see.

Photos lied for some time, and cameras make dreadful mistakes. But they're useful nonetheless.

Shari Blaukopf said...

You certainly don't need that James. I watched you turn the daytime street scene in Montreal's Chinatown into a night scene right before my eyes. That was magic.

Charley parker said...

An interesting aspect of the same branch of technology is "AI"-enhanced image interpolation. For better or worse, this promises to allow upsizing of images that were previously too small to be scaled without pixelation or blurring.

"Let's Enhance" allows 4 or 5 free test images before requiring purchase:

This article on Mashable gives an overview.