TFDS now supports the Croissant 🥐 format! Read the documentation to know more.

domainnet

Description:

The DomainNet dataset consists of images from six distinct domains, including photos (real), painting, clipart, quickdraw, infograph and sketch. Per domain there are 48K - 172K images (600K in total) categorized into 345 classes.

In this TFDS version of DomainNet the cleaned version is used.

Additional Documentation: Explore on Papers With Code
Homepage: http://ai.bu.edu/DomainNet/
Source code: tfds.image_classification.domainnet.Domainnet
Versions:
- 1.0.0 (default): Initial release.
Auto-cached (documentation): No
Feature structure:

FeaturesDict({
    'image': Image(shape=(None, None, 3), dtype=uint8),
    'label': ClassLabel(shape=(), dtype=int64, num_classes=345),
})

Feature documentation:

Feature	Class	Shape	Dtype
	FeaturesDict
image	Image	(None, None, 3)	uint8
label	ClassLabel		int64

Supervised keys (See as_supervised doc): ('image', 'label')
Citation:

@inproceedings{peng2019moment,
  title={Moment matching for multi-source domain adaptation},
  author={Peng, Xingchao and Bai, Qinxun and Xia, Xide and Huang, Zijun and Saenko, Kate and Wang, Bo},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision},
  pages={1406--1415},
  year={2019}
}