In the process of developing our image recognition technology we always try to test it in various environments. First there were book covers, posters and magazines. Then there were TV ads and multiple objects in one picture. And now it is time for beer, or more specifically, beer bottles.
Recently we released the beta version of shopping shelf recognition mode in our API at recognize.im. To give you an idea of how you can use this new technology we would like to show you results of one of our tests. We collected several different beer bottles, added their labels as reference images to our API and took some test pictures. You can see recognition results of one of these images above. It should be noted, that recognizing beer bottles is not an easy task – their labels are curved and often wrinkled or reflective. But out of all bottles in the image above, only few are left unrecognized, and all the recognized are recognized correctly.
To achieve this level of performance you need good reference images. In our test we compared two ways of obtaining pictures of beer labels. First, we peeled off labels and scanned them on the office scanner. You can see example in the image above (left label) – it’s not ideal, there are some visible wrinkles and reflections. In the second approach we took pictures of labels on bottles with compact camera (right label in the image above). Overall second approach yields better results, but we think that with better reference images (possibly graphic designs of labels) those results would be even higher.