PACS

Overview

PACS (Photo, Art painting, Cartoon, Sketch) is a domain generalization benchmark with 9,991 images of 7 object classes: dog, elephant, giraffe, guitar, horse, house and person. Every class appears in four visual domains, and the shift between domains is large: photos and art paintings are textured and colorful, while sketches are sparse black-and-white line drawings.

Domain	Images
photo	1,670
art_painting	2,048
cartoon	2,344
sketch	3,929
Total	9,991

Dataset Configurations

The original release comes with an official train/validation split for every domain (<domain>_train_kfold.txt and <domain>_crossval_kfold.txt, about 90/10), and its readme asks that this split be used for fair comparisons. The configs follow it:

all (default)

train and validation are the official splits of all four domains. There is no test split.

photo, art_painting, cartoon, sketch (leave-one-domain-out)

train and validation are the official splits of the other three domains, and test is every image of the named domain (<domain>_test_kfold.txt). This is the protocol of the original paper: the model never sees the test domain, and model selection uses the source-domain validation split only.

Data Structure

When accessing an example using ds[i], you will receive a dictionary with the following keys:

Key	Type	Description
image	PIL.Image.Image	227x227 RGB image (sketches are converted to RGB on a white background)
label	int	Class label (0-6)
domain	int	Domain index (0: photo, 1: art_painting, 2: cartoon, 3: sketch)

Usage Example

All domains

from stable_datasets.images import PACS

ds_train = PACS(split="train")
ds_val = PACS(split="validation")

sample = ds_train[0]
print(sample.keys())  # {"image", "label", "domain"}
print(ds_train.features["domain"].names[sample["domain"]])

ds_torch = ds_train.with_format("torch")

Leave-One-Domain-Out

from stable_datasets.images import PACS

# Train on photo, art_painting and cartoon; evaluate on sketch
ds_train = PACS(config_name="sketch", split="train")
ds_val = PACS(config_name="sketch", split="validation")
ds_test = PACS(config_name="sketch", split="test")
print(len(ds_test))  # 3929

References

• Project page and download: https://dali-dl.github.io/project_iccv2017.html

• Paper: https://arxiv.org/abs/1710.03077

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 = {Proceedings of the IEEE International Conference on Computer Vision (ICCV)},
  pages     = {5542--5550},
  year      = {2017}
}