cub-web

The CUB+Web dataset constructed by ourselves.

1. The training set consists of 750 annotated images (5 per class) from the 150 seen classes of the standard CUB dataset (Wah et al. 2011), along with 1205 unannotated web images obtained by Google with query 'North American Bird';
2. The test set consists of 2946 unannotated images from the 50 unseen classes of CUB.