Visual Distance Measures for Object Retrieval

This paper describes an enhanced visual distance measure for image features, and evaluates its effect on object retrieval accuracy for several standard datasets. The measure incorporates semantic proximity information that is automatically extracted from each dataset in an offline step. It is designed to overcome errors introduced by feature detection and quantization in the "bag-of-words" model. We define a cross-word image similarity measure using this visual word distance, and show that it improves object retrieval precision for several datasets. It involves minimal additional query time cost, and can be embedded into any object retrieval method that uses a "bag-of-words" model.