Plenty of companies are taking advantage of machine learning to tag and search visual content. Pinterest lets you find visually-similar images in order to track down that recipe or jacket you’re looking for, and Pornhub is using machine learning to automatically identify porn stars in videos. Stock image company Shutterstock, though, has developed one of the more novel implementations of this sort of technology: using machine learning to identify the layout of images.
The new tool, launched today, is currently available on the company’s test site, Shutterstock Labs. You can search for various elements (in the case of the link above, wine and cheese) and then move icons about to specify where you want them to appear in the image. You can also specify where you want blank space — perfect for designers who are looking for a specific layout and a place to put text.
Based on experiences with this beta, the tool is quick and well-designed, but not always useful. Finding pictures of wine and cheese positioned in various ways on a table is fine, but there’s no way to specify more complex compositions, including any arrangement involving 3D space. Also, moving the icons just a touch to the left or right sometimes returns very different results.
Interestingly, Shutterstock notes in its press release that it hadn’t intended to train algorithms to categorize spatial arrangements, only to recognize objects. The information about layout was just scooped up as part of the analysis, and later identified as a useful category of information. It’s a good example of how this sort of machine learning can deliver improvements for users that aren’t very exciting, but are, at least, pretty useful.