June 3, 2015 weblog
Google's Im2Calories calculates calories in a meal by examining a photo
At the Rework Deep Learning Summit meeting last week, Google researcher Kevin Murphy unveiled a few details of a project he and coworkers have been working on—called Im2Calories, it is an app that allows a user to snap a photo of a plate full of food with their phone and get back a calorie estimate.
Such an app would obviously have to rely on multiple types of technology—image recognition, access to a food/calorie database and perhaps most importantly, an ability to learn. Murphy readily acknowledged that the app in its current format is not all that accurate when first used, but as it learns over time, it is hoped that it will get better at identifying what is on a plate—for example, noting details in a photograph that help distinguish a fried burger from one that is grilled—one has a lot more calories than the other.
The project is part of a "deep learning" effort going on at Google and other companies—where AI technology is used to help marry image and pattern recognition—the deep learning part comes as the system learns, and unfortunately, because this is all still in its infancy, that means via user training. With Im2Calories, for example after a picture is taken and the system analyzes what it sees, it spits out a report listing what it believes it has found: poached eggs, buttered toast, two strips of bacon, etc. At that point, the user has the option to make corrections, telling the system that in fact, the eggs were boiled, or fried, or whatever. The system would then make that correction, and add a note to its database so that it would not make the same mistake in the future.
Presumably, such an app would undergo some initial training before being sent out into the populace, though in this case it is not clear if that will happen at all with Im2Calories. Murphy said the app was being developed as a research project, which means that at this time there are no plans to actually release it. That could change of course if the app proves highly accurate and if users hear of it and beg for Google to give it to them.
More information: via Popsci
© 2015 Tech Xplore
