14 comments

[ 3.7 ms ] story [ 36.4 ms ] thread
I call bullshit. How would it be possible? I have two bets how this can remotely work:

1. They use some newly discovered, color dependent property of original negatives, but the following quote debunkes that:

> If you have a ready-made files of black and white image or video, color recovery process will be much faster.

2. They use machine learning.

Well if they use machine learning then they just teach the computer how to do colorization and produce plausible results. It's not retrieving the original colors. It could be still a good idea though.

The people over on reddit's colorizedhistory subreddit do such a marvelous job. If it were applied to film, it would be fantastic. Of course, it's too hard to do every frame. But perhaps an artist could color one frame of each scene, and software + machine learning can deduce the coloring of the rest of the frames of that scene.
You could do keyframes and then in-beteween. If it is the same scene then use the previous keyframe as a guide. That should speed it up considerably.
I suppose the process would be analogous to and as labor intensive as the recent conversions of older movies to 3D, i.e. practical for many movies.
Anyone who has grown up with b&w television will tell you that you were actually able to tell the colors of objects, when you were used to it.

The other way round, if you happen to have done any serious recoloring work (e.g., those car configuraters found at the web[1]), you probably know that it's not just about changing the hue. The various reflective properties, diffusion and response to exposure are quite different for different colors and it boils down to reworking the entire histogram to achieve some credible appearance. So, yes, I think it's quite plausible that you may analyse the properties of a surface captured in b&w and recover the original color properties with some limited precision.

(Now, that we're spoiled by highly saturated displays with a limited color range and are used to ignore artifacts of image compression, this may seem quite far fetched. But when you are used to deal with analog images, yes, it's plausible. P.S.: And while we are at it, please stop colorizing images with colors that are obviously not the originals. It's not all the same, just because it comes in b&w.)

[1] Before there were renderings, you would get a product shot in a single color and all the other color versions had to be derived from this.

It'd be nice to have a description of how it works, but the site's explanation is vague to the point of nonexistence.
Walter if they explained how it worked then the gullible wouldn't buy.
You have so little faith in the power of gullibility [1]

[1] http://www.yoyodyneindustries.com/inductions/beacon_inductio...

I don't actually believe that Carole Lombard was quite that pink in real life.
Wait, how did they find the original color of a fake picture (the giant grasshopper, or "gracehopper" as they call)?
It's hard to find an actual explanation for the process on the site, after a bit of rummaging this is the closest I could get: http://i.imgur.com/YsbKHg1.gif

Still, it's a surprise to discover that changing a photo to black and white is a lossless operation that can be reversed.