FeedbackBypass is a technique able to improve the overall performance of image similarity searches by learning user preferences that determine the similarity criterion and providing adequate approximate answers to trade-off the quality of results for execution speed. The basic idea consists in properly maintaining user preferences (i.e. relevance feedback) from user interactions (i.e. feedback loops) to either "bypass" the feedback loop completely for already-seen queries, or to "predict" near-optimal similarity criterion for new queries.
Example: The top line in figure shows the 5 best matches computed using the default similarity criterion for the query image of type "animal". The bottom line shows the results obtained for the same query when the similarity criterion suggested by FeedbackBypass is used.
Two different implementations of FeedbackBypass have been developed: A wavelet-based version that defines a structure called "Simplex Tree", and a second one that uses Support Vector Machines (SVM).