SHARE

A few years ago we told you about a new technology called Algorithmia that could automatically guess accurate colors for black and white images. While the program was cool, the results were often far from perfect. A new tool called Colourise.sg is now available and so far it appears to be more accurate that the Algorithmia program.

ian jumping in black and white
The original image shot on Kodak Tri-X black-and-white film. Stan Horaczek

Colourise.sg was developed in Singapore by a company called GovTech during a January hackathon. According to the creators of the project Algorithmia produced colorized image that lacked intensity, which makes the final product appear inauthentic. The goal of Colourise was to create a program that was specifically trained to colorize old Singaporean photos. The creators did this by training the tool with historical images from Singapore. In contrast, Algorithmia was trained using 1.3 million images from ImageNet, a database often used by researchers at Stanford and Princeton specifically to train AI.

ian jumping image with color filter
The colorized version of the photo as interpreted by Algorithmia. Stan Horaczek

Although the original tool did an okay job differentiating humans from nature, it often applied color with very rough borders. It also seemed to do better with human subjects when they appeared quite large in the frame. We were curious to see how Colourise.sg would interpret these same images.

ian jumping image with colourised algorithm
The colorized version as interpreted by Colourise.sg. As you can see the new program did a nice job of filling the entire image with color. The desaturated colors make it look a bit like it was shot with an old roll of color film. Stan Horaczek

Colourise.sg is a lot better at filling in an entire image with color—you won’t find any rough borders in the image output. The colors tend to be desaturated, which makes the images look like they were shot on expired film stocks. The images processed with Colourise also tend to have a cooler tone than the ones processed by Algorithmia.

RELATED: You should develop your own black-and-white film. Here’s how.

yoga in nature in black and white
The original black and white image fed into the site. It was converted (using a very basic black-and-white conversion) in Lightroom. Stan Horaczek
yoga nature image colorized
How Algorithmia interpreted the colors in the scene. Stan Horaczek

Colourise also seems to do a better job at identifying human subjects even if they are small in the frame. When we sent a black and white version of the above photo through Algorithmia it totally missed the person doing yoga and colored her the same gray as the rocks.

Yoga in nature photo colorized
Colourise did a better job than Algorithmia at identifying the woman doing yoga in the bottom of the frame. Stan Horaczek

In the Colourise version her skin tone does appear to be a little green, but she is being processed in distinct colors from her background. Check out the original digital color file below for a comparison of how it did.

yoga in nature photo in color
The original color photo for comparison, shot on a Canon EOS 5D Mark III. Stan Horaczek

It’s a cool tool and its interesting to see how this AI tech has advanced in just a few years. While tools like this will be no means replace the important work being done by professionals, it’s neat to see it quickly analyze these scenes and what they might have looked like if they were shot in color. Check it out for yourself.