Putting Robots to Work on the Past

In 1896, Frenchman Charles Moisson, one of the few cameramen working for Lumiere company, was in Moscow to attend the coronation of Russia's last tsar, Nicholas II. He made a widely-known short clip of bustling Tverskaya street, showing the corner of the still-standing National Hotel and passing walkers and carriages.

Now Russian techie Denis Shiryaev [about] has put Moisson's camera work through several neural networks, to enhance and colorize the original monochrome clip. Internet users have praised his work as an incredible "time machine" that gives us a glimpse of the past that looks more realistic than ever.

We asked Shiryaev about his interest in using the work of robots to produce this nostalgia-laced result. For more technical information about his work, click on the clip to read the YouTube description in Russian and English.

RL: In a nutshell, how does this work? Do you upload the video into a program? What else does it need to know about the video to get the colors right?

DS: I put the video through an ensemble of neural networks, each of which plays a role. One cleans the video of noise, another creates additional frames to make the video play more smoothly, a third one increases the resolution, and finally one "colors" the video. Each algorithm has been honed on its own set of data, for example, the coloring neural network knows that people's faces are usually beige, grass is green and the sky is blue, because that's how the neural network was trained. 

RL: Do you think the result would be better if the video was colored by hand?

DS: Robots and algorithms are cheaper and faster than human work. Restoration can definitely be done by people, but with each year, work by algorithms will improve. People historically lose the competition to robots, and in five years' time neural networks will not only improve data that exists in an image, but also add additional faces, objects, and anything else. In fact, restoration is basically not done purely by hand anymore, because restorers use a lot of computer programs to enhance the image.

RL: Why do you choose clips like these to work with? Are you nostalgic for the era before the internet?

DS: I like history and that the image comes alive when you put it through the different algorithms. I choose the clips based on my own taste and the advice by YouTube commentators, who suggest other ideas for restoration.

RL: Reading the comments, especially to the clip "Arrival of the Train", it seems that a lot of people are struck by the thought that "everyone in the video is now dead" even the little children. Isn't it strange that a restored video from 1896 draws this response? 

DS: Old clips that we usually see look different from videos that we're normally used to seeing on the internet. People on them are usually seen running to the sound of music or the projector noise, so the videos look "ancient" and you don't associate yourself with the people in them. When I restore the clips, people come alive and the clip stops seeming artificial. You can no longer get rid of the feeling that all of those people are real.