In this paper the authors experiment with 10,000 image classes based on ImageNet. As ImageNet is based on Wordnet, they have a semantic tree of the categories.
It should be noted that this paper is from 2010. Hence before AlexNet. They don't use CNNs in this paper.
## Key findings
* A relationship between visual similarity and semantic similarity exists
* Classification can be improved by exploiting semantic hierarchy
* Computational difficulties with 10,000 classes
* More classes -> lower mean accuracy
## See also
* [What makes ImageNet good for transfer learning?](https://arxiv.org/abs/1608.08614) ([slides](https://www.dropbox.com/s/vfmncjnyh57glkc/NIPS_LSCVS_ImageNet%20Analysis.pdf?dl=0))