fa2431bf9d65058fe34e9713e32d60e6-AuthorFeedback.pdf

InR2'swords,17 "the employment of these ideas together for attribute localization and ZSL is quite interesting and seems to lead to18 consistentgoodperformance"(seeTab. In contrast, our APN21 aims to improve the image representation for zero-shot learning by learning prototypes that predict attributes from22 intermediate features. The evaluation is conducted on the image-level attributes of test images on CUB31 dataset. Foreach attribute, we learn aprototype on intermediate CNN features to34 regress the desired attribute. So it is expected to learn specific attributes rather than just the color or parts, which is35 empirically confirmed by the attribute prediction accuracy which is higher than the attribute prediction model (See36 previous response toR2), andthequalitativeresults inFigure.