Really, it's amazing what they're doing these days with images. Automated object identifiers can scan a photo of someone playing tennis and label the objects in the picture: Person, tennis racket, lemon, tennis court ...
Hold up. Lemon? Yep. Lemon.
And therein lies the problem. Any person would look at that list of tags and know something is amiss. One of these things just doesn't belong. Either it's a doctored photo or the tag is incorrect.
Until now, it would take human intervention to fix the problem. But yesterday, computer scientists from UC San Diego and UCLA shared how they've managed to add good old fashioned common sense to automated image labeling systems.
And, surprisingly, the key to the whole problem is a Google Labs widget called Google Sets.
According to vnunet, the researchers -- or to use the British slang, "boffins" -- built a system that uses three steps:
There's a little more to it than that, including that the researchers ran their object categorization model on the segments, and not pixel-by-pixel. According to Science Daily, the process with Google Sets increased the average categorization accuracy more than 10 percent for one dataset and 2 percent on the second dataset.
The paper, "Objects in Context," by Andrew Rabinovich, Carolina Galleguillos, Eric Wiewiora and Serge Belongie, was presented Oct. 18 at the IEEE International Conference on Computer Vision in Rio de Janeiro, Brazil.