In this article, I will compare the automatic colourising services colorize-it.com, colourise.sg and auto-colorizer.com to hopefully offer you some perspective of how useful they are, individually and collectively.
I will always present the original colour image after the various auto-colourising outcomes, not necessarily in expectation of getting the exact same result back but at least to offer some perspective against which to compare the results.
NB auto-colorizer produces only a small “free” output image, which is used throughout.
Sample 1 (landscape, dawn/dusk)
Image taken from promotional material for the upcoming Joker movie. The intention was to use material that the algorithms would have been unlikely to be trained on.
My take: Auto-colorizer seems to have an inkling that there’s a colour gradient. For colourise.sg, skies seem to be blue, no questions asked. To be fair, whether the colour gradiant is real of added in post-production could be up for debate. It’s a close call, but of the three colourisations, the colourise.sg image would probably most convince me that it’s real.
Sample 2 (night street scene with neon displays)
My take: I found it interesting that none of the algorithms are sufficiently familiar with traffic lights to know that when there are three round things one above the other, the bottom one is green and so on. Also, none of them seem to put ANY hue differentiation into the output. You’re basically getting a monochrome image back – just with a colour cast and in one case, even that is faint.
Sample 3 (indoor portrait, Caucasian female)
The subject here is the artist Anne Maria Udsen via Wikimedia Commons.
It’s very clear that the Singaporean effort was more carefully trained on images of people.
My take: At this point, it seems that colorize-it hardly works at all. It just makes everything red to varying degrees. Colourise.sg produced a startlingly good image for once.
Sample 4 (clown stage portrait)
This is a portrait of Lasse Beischer as a clown. Source.
My take: Colorize-it seems to have landed a lucky punch for once, almost as though this or a similar image was in the training set. auto-colorizer did not produce output – it seems the clown broke the machine. I tried several times, but no luck with this image.
Sample 5 (daytime landscape with rainbow)
My take: Quite a mixed bag. Auto-colorizer produced the most unappetising vegetation of the three. None of them correctly recover the rainbow.
Sample 6 (outdoor movie scene, colourful clothing)
My take: It looks like colourise.sg takes the approach of not colouring areas at all if it’s not sure about them, and auto-colorizer may be doing something similar. Even colorize-it left an uncoloured patch. As for good news, all recovered the dark skin tone at least somewhat correctly, to varying degrees of interpretation.
Sample 7 (iconic movie shot, strong colour)
My take: In a nutshell, sky and skintones are reliable, but nothing else is. auto-colorizer seems to be the only one to correctly identify the blurred vegetation. On the other hand, colorize-it produced some more interesting colour, which I found engaging in spite of occasional imperfections, especially to do with edge recognition. Still, in this task, it’s my personal favourite out of the three.
And on to part 2…
Sorry for making this a cliffhanger, but since this is quite a long read already, I’ll give you my conclusions in part 2.