Flickr AI Project

I wonder if anyone is using the flickr photo sharing database to train AI. It would be an amazing task to get an AI to play the flickr game fastr.

fastr selects a keyword then searches the flickr database using that keyword. It retrieves 10 images tagged with the keyword. One image at a time is exposed to the players. If a player guesses the keyword having been shown only one picture they get ten points. After two images they receive 9 points and so on down the line.

The game illuminates the difficulty of the image classification problem in computer vision. Maybe it's because I watched Robosapiens last night, but the game really got me thinking about robot vision.

Take this image as an example. It is tagged with dog, kyla, 2006, and Boston terrier. Dog and Boston terrier are derivable from the photo, but kyla and 2006 are not. So, the flickr database is probably not optimal for training image classification algorithms.

What can you tell from the photo: Dog, Chair, Boston Terrier, Fiberglas Chair, Orange Chair, indoor, hallway, shadow of photographer, subject illuminated from behind camera, dog seated, and dog alive. There is probably a bunch more information one can derive from the photo, but I'll stop.

A more useful database would be created if you could classify the image by indicating which parts of the image allow you to derive which information. Meaning maybe you could outline areas of the image and tag those areas. For example outline the dog and tag it: dog, dog seated, dog alive, and Boston terrier. Then outline the shadows and tag it: subject illuminated from behind camera.

Anyway, I find the whole problem intriguing.

Update: I guess there are some individuals that can derive the information kyla and 2006 from the photo. It's all a matter of perspective. For a personal robot that info would be relevant but for a generic system, not so much.

No comments: