Press "Show random images" to randomly browse the database. Select images as positive "+" or negative "-" examples in your query. Press "Perform query" to retrieve images matching your query.

Overview

This system uses a new algorithm which approximates the perceived visual similarity between images. The images are initially transformed into a feature space which captures visual structure, texture and color using a tree of filters. Similarity is then measured as the distance in this perceptual feature space

A typical query consists of a small set of images which are representative of a broader class (e.g. images of automobiles or images of city skylines). From the example images a characteristic signature in feature space is computed and is compared to the features of each image in the database. The closest database images are returned.

We hypothesize that there is in fact no clear distinction between local texture and global structure: that they are simply two ends of a continuum. Our algorithm represents images at many levels of resolution: measuring color, edge orientation, and other local properties at each resolution. The visual properties captured by these local operations changes at different scales. The goal of our approach is to pay attention both to local texture and global structure. Thus by finding patterns between image regions with particular local structural organization, more complex -- and therefore more discriminating -- features can be extracted.

For more information you can view an online abstract or see slides which describe the process in detail.

Example retrieval results are also available.