Facebook's Image Recognition has Been Overhauled to Insane New Proportions

Mashable

When someone opens their announcement with the claim that internet searching has been revolutionised, ears tend to prick up. In a blog post earlier this week, Facebook's Director of Applied Machine Learning (yep, that's a thing) Joaquin Candela boldly stated that the days of searching for images based tagging and captions were over.

According to Candela, Facebook have developed a new set of machine learning based tools which will allow you to search for images under parameters like, say, 'red shirt', 'driving', 'dancing' or 'sunshine', and regardless of how the images are tagged, the system will know what you mean.

Facebook have been playing around with this tech for a while, initially billing it as a way to help the visually impaired understand what's in an image. The actual platform for this is called Lumos, and it basically just scans images pixel by pixel, to the point where it can recognise specific visual features automatically. As well as allowing for a deeper form of searching, it will also help Facebook improve the category tagging it uses for images.

If this system continues to advance at the same rate (which it almost certainly will), then the possibilities are vast. Every time you do an image search, there are thousands, perhaps even millions of images likely ideally suited to your search terms that you just won't see. Tagging and captions only get you so far, and there are reams of images all over the internet which have neither.