CloudCutout Applies Machine Learning To Image Editing

Danish startup CloudCutout is applying machine learning techniques to the quotidian yet pernickety process of isolating an image from a background. Their initial target market is school photography, with the aim of undercutting the market by charging half the “standard knockout” bulk cost (50c) per image.

The Copenhagen-based team says its algorithm can automatically cut out studio images from consistent backgrounds, such as green- or blue-screens, but also from “standard backgrounds” used for school photographs, and also from the white backgrounds popular for product shots — saving time and money. Or that’s the pitch.

The underlying algorithms are apparently based on co-founder Toke Jansen’s machine learning PhD, and have been trained to recognize cutout patterns by being exposed to 100,000+ professional cutouts. The startup was founded early last year, funded by the other co-founder, Henrik Paltoft’s, cutout agency. It’s also now taken in a $500,000 seed round from Danish based VC firm Seed Capital.

“To provide high quality cutouts, the core of our engine exploits recent advances in spectral graph theory and neural networks. The computation of pixel transparencies (the alpha channel) for a single image involves solving multiple large-scale equation systems, as well as carrying out multiple feed-forward passes in our neural networks,” say the founders.

“We pose the problem of determining an alpha channel of an image as a machine learning task. Compared to usual chroma keying, this allows us to consider a much broader range of backgrounds since the model will learn, i.e., texture representations from existing training data.”

They’re launching a cutouts for schools service this month as their first offering, initially limited to green screen knockouts, before they expand out to support other backgrounds. Users upload their photos to CloudCutout’s platform for processing, and get their cutout images returned within 48 hours. The cost per image is 25 cents, although there’s a minimum requirement of 50,000 images per year so that’s only going to make sense for larger schools. Unsurprisingly, their primary focus is the U.S. schools market.

They also plan to launch a Photoshop plug-in later this year to enable users to manually tweak shots that still require adjustment after being put through their algorithm.

“This fall, we are working with chosen partners within school photography and pack shots. Later this year, we are planning to launch a more generic Photoshop plug-in for the vast numbers of graphic designers worldwide,” say the founders. “The business model for the service will be a certain price per image, and for the plug-in it is probably going to be a freemium model, with subscription opportunities with price depending on quantity.”

What percentage of CloudCutout cutouts will require manual tweaks? None of the initial green screen knockouts, according to the founders. Generally speaking, they add that if the cutout subject is a product or person that’s “distinctively different” from the background then the vast majority (90 per cent+) won’t require additional tweaks.

“We help cutting out all images,” they add. “Some perfectly the first time, some will have a bit of retouching to be done, but in these cases, our Photoshop Plugin will make the designer able to continue the work in the industry’s standard tool and thereby significantly shorten the time of cutting out any image.”

To ensure they can scale to accommodate demand for the cutout service they’re distributing their cutout engine on IBM SoftLayer and Amazon Web Services. “Computations related to neural networks are executed on bare metal SoftLayer nodes, that provide the most recent NVIDIA Tesla K80 GPUs, whereas remaining parts of our infrastructure are hosted on AWS,” they note.