I'm building an application which needs to take an image and infer tags related to it. Those tags can be about things, adjectives or even emotions related to the picture.
I've already found ALIPR. But I tested it, some other people tested it also and it doesn't perform well. ALIPR makes too many mistakes in the set of 15 predicted tags. At least for my application, it is better to have few but correct tags.
Preferably, the API should be web-based and free. Any suggestions?
Thanks in advance!
I think if images could be labeled automatically, Google would have abandoned the image labeler a long time ago. Unfortunately, computers have a lot of trouble understanding images.
Edit:
If you are interested in computer vision research have a look at CVPapers, especially Open Source Computer Vision Implementations. Automatic image labeling is far from being solved (unless you have a very specific/restricted set of topics).
Quote from The Google Guide from Tuesday March 13, 2007:
The words “Larry Page” and “Sergey Brin” appear near images of Eric Schmidt, or in image captions, or in links to those images. Google makes a guess that the words are related to the image. Google technology isn’t yet to the point where it can tell what’s in an image by looking at it directly.
HTH, don't get your hopes up too high.
P.S.: I hope you (or someone else) proves me wrong and shares it here with me ;-)
Edit2:
I just stumbled across the Voc 2010 Challenge, which, in my opinion, illustrates very well the current state of computer vision advances. In one of the challenges the contestants have to find an object (from a very limited set of objects) in the image and classify it. On the result page you can see, that one of the algorithms manages to classify air plane with a 93% accuracy, but "fails" at other categories.
That is just for the quest to find the "things", not even adjectives or emotions.