Automatic Matching Tool Selection Using Relevance Feedback in Mars

For a given visual feature, due to the diversity of human's subjective judgment, a visual information retrieval system that supports a single pre xed similarity measure will result in poor retrieval performance. To address this problem, this paper proposes the concept of similarity matching toolkit which consists of different similarity measures simulating human's perceptions of the given feature from di erent aspects. The toolkit supports a feedback-driven tool selection mechanism which adapts to the similarity measure that best ts the user's perception. To illustrate the advantage of the proposed toolkit approach, we apply it to shape-based image retrieval. The paper describes a shape matching toolkit consisting of four transformation-invariant and computationally e cient matching tools and describes how relevance feedback can be used for automatic tool selection. Experimental results validate the exibility of the matching toolkit and show the e ectiveness of the relevance feedback for shape matching tool selection.