A New Adventure

Image Retrieval - CIVR 2006

Right now I am sitting in a chair which is in a conference room in Mission Palms Hotel, in Tempe(Phoenix), Arizona, attending Conference on Image and Video Retrieval 2006. The task is what google has done to text, easily and efficiently retrieve images from a database. Lots of research is going on at different levels.

First the image is needs to be represented in computer, I know we have the color data for every pixel but thats not enough/rich enough to perform retrieval, so one chunk to people are working on what would be the best way to represent image so that it stores the image features efficiently.

The second step is to find which image features to save or what are the 'salient' points or 'interesting' points in a image that one would use to retrieve images. The basic is image color, image texture to little more complex as shape of objects. The notion is that both image representation and saliency should be able to 'generalize' the concepts one would be interested in.

The third direction of research is once we have the representation of the 'interesting' points, how do we match the given representations of two images to say wether they are similar or not. Two points in the space can have different measures of distance between them, the most common being a eucledian distance or an L2. Based on the feature space chosen, one can from very many metrics.

Talking at higher levels, one has then the notion of extracing image features of a concept that can be visually very different. Most commonly cars can have very many colors, shapes, direction of view, illumination conditions, backdrops they are while ckicking the photograph. Is is possible to still learn a mathematical model of a general term car? This is semantic image retrieval which is best with the human notions of similarity as we very easily can abstract out the concept of car from any image of car in various conditions.

The destination is far far away, I am not even sure that it can be solved in full generality, but history tells us that future is not deterministic. It is random with probabilites of unexpected events having spikes. Or is it destined? ;-P

posted by Nikhil @ 3:59 PM