Adapting k-d Trees to Visual Retrieval

The most frequently occurring problem in image retrieval is find-the-similar-image, which in general is finding the nearest neighbor. From the literature, it is well known that k-d trees are efficient methods of finding nearest neighbors in high dimensional spaces. In this paper we survey the relevant k-d tree literature, and adapt the most promising solution to the problem of image retrieval by finding the best parameters for the bucket size and threshold. We also test the system on the Corel Studio photo database of 18,724 images and measure the user response times and retrieval accuracy.