Domain Generalization (PACS)

This is the track of the state-of-the-art methods on PACS benchmark.

Domain generalization (DG) assumes a model is trained from multiple observed domains while it is expected to perform well on any unseen domains. PACS consists of Art painting, Cartoon, Photo and Sketch domains, which so far considers the largest domain shift as it is from the different image style depictions.

Dataset

Name Year Domains Images Classes
PACS 2017 4 9991 7

Benchmarks

Base Model Methods Venue Art painting Cartoon Photo Sketch Ave
AlexNet Li et al. ICCV17 62.86 66.97 89.50 57.51 69.21
AlexNet Li et al. AAAI18 66.23 66.88 88.00 58.96 70.01
AlexNet Massimiliano et al. ICIP18 64.1 66.8 90.2 60.1 70.3
AlexNet anonymous Under review in ICLR19 68.1 78.6 92.1 66.1 76.2
AlexNet anonymous Under review in ICLR19 66.8 69.7 87.9 56.3 70.2
AlexNet anonymous Under review in ICLR19 61.2 63.8 82.9 59.0 66.7

References

[1] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In ICCV, 2017.

[2] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Learning to generalize: Meta-learning for domain generalization. In AAAI, 2018.

[3] Massimiliano Mancini, Samuel Rota Bulo`, Barbara Caputo, Elisa Ricci. Best sources forward: domain generalization through source-specific nets. In ICIP, 2018.

Maintainers

Yongxin Yang and Da Li