Someone Used Neural Networks to Add Color to This Enhanced 1896 Footage

What’s better than upscaling a film from 1896 to 4K 60 FPS quality? Taking the next logical step and adding color, of course. Following up on the work of Denis Shiryaev, another YouTuber added color to the film L’arrivée d’ un train en gare de La Ciotat, for all of us to enjoy. It’s surprisingly successful, with just a few glaring mistakes.

Neural Networks can be a great way to get complicated work done quickly. But, training a neural network doesn’t involve telling it every single thing to do. Instead, it’s often a case of working backward.

If you want to train a neural network to upscale an image, you start with high-resolution photos and reduce their quality. The downsampled pictures become a starting point for neural networks to learn from, and the original high-quality images become the “right answer.”

When Denis Shiryaev used several tools to upscale L’arrivée d’ un train en gare de La Ciotat, they took guesses on how to improve the video, add new frames, and smooth everything out. The results were impressive.

Now, another YouTuber (Deoldify videos) took that footage and used DeOldify tools to add color. DeOldify works like many neural network tools, but with an added twist. DeOldify relies on two neural networks, one that adds color to the video and another that criticizes the result.

One neural network teaches the other to work better and produce more realistic results. The final result is a reasonably convincing color version of L’arrivée d’ un train en gare de La Ciotat.

Train enthusiasts, however, will be quick to point out that the A.I. got important details wrong. The train seen in the video should feature green livery, as locomotives owned by the PLM railroad always featured those colors at the time. Look closely at the reflections coming off the train, and you’ll see spots that weren’t adequately colored either.

Still, the results are impressive, especially when you stop to think that a human didn’t do any of the work. With time and effort, the results and accuracy will improve.

 

Josh Hendrickson Josh Hendrickson
Josh Hendrickson has worked in IT for nearly a decade, including four years spent repairing and servicing computers for Microsoft. He’s also a smarthome enthusiast who built his own smart mirror with just a frame, some electronics, a Raspberry Pi, and open-source code. Read Full Bio »

The above article may contain affiliate links, which help support Review Geek.