Abstract

Recognizing fifine-grained sub-categories such as birds and dogs is extremely challenging due to the highly localized and subtle difffferences in some specifific parts. Most previous works rely on object / part level annotations to build part-based representation, which is demanding in practical applications. This paper proposes an automatic fifinegrained recognition approach which is free of any object / part annotation at both training and testing stages. Our method explores a unifified framework based on two steps of deep fifilter response picking. The fifirst picking step is to fifind distinctive fifilters which respond to specifific patterns signififi- cantly and consistently, and learn a set of part detectors via iteratively alternating between new positive sample mining and part model retraining. The second picking step is to pool deep fifilter responses via spatially weighted combination of Fisher Vectors. We conditionally pick deep fifilter responses to encode them into the fifinal representation, which considers the importance of fifilter responses themselves. Integrating all these techniques produces a much more powerful framework, and experiments conducted on CUB-200- 2011 and Stanford Dogs demonstrate the superiority of our proposed algorithm over the existing methods