Well, technically you're right, but I don't think the pigments in our skin made huge evolutionary advancements in ~100 years. I'd guess people were working more outdoors, but that's it.

So if you were to recolor a film you can take skin colors from today. But an advert, car color or fabric is incredibly difficult, even if you have the original, since the material will have degraded. Even if you repaint it with the original paint it will look different, since the paint will have changed, or if you use new paint the process to manufacture the paint (->led).

It might as well be another universe.

which skin colour is a question here as well. There are a lot of variations and I'm not sure that a black and white film is going to contain enough information for you to know. There's a danger of white-washing (or the reverse!) history by mistake.

true - but we could just recolour it if needed

I think the point is that you simply don’t know in a lot of these videos. They’re random bystanders, not known figures with photographic evidence. So you’re not going to know what to recolour it to, or if it’s even wrong in the first place.

Not all shades of skins are the same within one specific demographic.

Unless there's also a way to accurately extract the specific hue, you will never get anything accurate.

Anyone that ever wore foundation can tell you that human skin has tones, undertones, different reflective qualities, etc.

Just look at https://www.temptalia.com/wp-content/uploads/2018/05/summer-... or https://cdn.shopify.com/s/files/1/0009/0000/5932/files/Shade... or this https://pbs.twimg.com/media/B_zp3U1XAAAk4Iz?format=jpg&name=...

A couple years ago I used deep learning to recolor some old family photos. The net I was using would do it's best guess first, then present a limited set of options for me to choose from when it didn't know what to do. In general, its first pass would immediately make human faces look correct, but leave everything else really desaturated.

I think some of it comes down to the fact that convnets are really good at recognizing human "shapes" (faces, silhouettes, etc) and also the valid color space for human skin tone is pretty constrained if the photo was well lit and you have the luminance value for a grayscale image.

The network was particularly bad at getting the colors of balloons in a picture. I don't know what the real colors were, but the network was trying to tell me 4 of the balloons in a picture were pale yellow and beige while one was a super saturated red.