Abstract

We address the problem of large-scale fine-grained vi-sual categorization, describing new methods we have usedto produce an online field guide to 500 North American birdspecies. We focus on the challenges raised when such a sys-tem is asked to distinguish between highly similar species ofbirds. First, we introduce one-vs-most classifiers. By elim-inating highly similar species during training, these classi-fiers achieve more accurate and intuitive results than com-mon one-vs-all classifiers. Second, we show how to esti-mate spatio-temporal class priors from observations thatare sampled at irregular and biased locations. We showhow these priors can be used to significantly improve per-formance. We then show state-of-the-art recognition per-formance on a new, large dataset that we make publiclyavailable. These recognition methods are integrated intothe online field guide, which is also publicly available.