What makes ImageNet good for transfer learning?

Minyoung Huh, Pulkit Agrawal, Alexei A. Efros

Published 2016-08-30Version 1

The tremendous success of features learnt using the ImageNet classification task on a wide range of transfer tasks begs the question: what are the intrinsic properties of the ImageNet dataset that are critical for learning good, general-purpose features? This work provides an empirical investigation of various facets of this question: Is more pre-training data always better? How does feature quality depend on the number of training examples per class? Does adding more object classes improve performance? For the same data budget, how should the data be split into classes? Is fine-grained recognition necessary for learning good features? Given the same number of training classes, is it better to have coarse classes or fine-grained classes? Which is better: more classes or more examples per class?