PACS Dataset

Overview

PACS is a benchmark dataset for domain generalization in image classification, introduced in "Deeper, Broader and Artier Domain Generalization" (Li et al., ICCV 2017).

It contains 9,991 images across 4 domains and 7 object categories, with significantly larger domain shift than prior benchmarks like VLCS — averaging a 20.2% cross-domain performance drop versus 10.0% for VLCS.

Domains

Domain	Description
P — Photo	Real photographs
A — Art Painting	Artistic paintings
C — Cartoon	Cartoon-style illustrations
S — Sketch	Hand-drawn sketches

Classes

7 categories: dog, elephant, giraffe, guitar, horse, house, person

Dataset Statistics

Domain	Images
Photo	~1,670
Art Painting	~2,048
Cartoon	~2,344
Sketch	~3,929
Total	9,991

Usage

The standard evaluation protocol is leave-one-domain-out: train on 3 domains, test on the held-out domain. This yields 4 cross-domain tasks:

• Train on A, C, S → Test on P
• Train on P, C, S → Test on A
• Train on P, A, S → Test on C
• Train on P, A, C → Test on S

Citation

@inproceedings{li2017deeper,
  title={Deeper, Broader and Artier Domain Generalization},
  author={Li, Da and Yang, Yongxin and Song, Yi-Zhe and Hospedales, Timothy M},
  booktitle={ICCV},
  year={2017}
}

Uploaded By

Mohammed Azeez Khan — used for domain generalization experiments at Carnegie Mellon University (EEG P300, motor imagery, fMRI neuroimaging).