Addressing Partial Relevance in Image Retrieval through Aspect-based Relevance Learning

In this paper we focus on a number of issues regarding special structure in the relevance feedback learning problem, most notably the effects of image selection based on partial relevance on the clustering behavior of examples. We propose a simple scheme, aspect-based image search, which directly addresses these issues. The scheme additionally allows for natural simulation of the relevance feedback process. By means of simulation we analyze retrieval performance, sensitivity to feature errors, and demonstrate the value of taking into account partial relevance for a database of decoration designs.