Photo to Rest Generalization
3 papers with code • 2 benchmarks • 1 datasets
It is the practical scenario of training on set of easy-to-collect real photographs and evaluate on the rest of diverse-styled domains (art, cartoon, sketch). Photo-to-rest generalization is a special case of the single-source domain generalization (SSDG) task.
Using only real photographs for training is the only way for SSDG to be compatible with ImageNet pre-trained networks. For any other domain as source, access to the photo domain through ImageNet pre-training violates the hypothesis of the task of a single domain seen during training, taking the task closer to the multy-source domain generalization (MSDG). Finaly pre-training on imagenet, training on a non real photograph domain and testing on real photographs violates the SSDG hypothesis that the target domain should not be used during training.
A large part of the generalization community chooses only this task over SSDG for pre-trained networks and for datasets that contain the photo domain.
Benchmarks
These leaderboards are used to track progress in Photo to Rest Generalization
| Trend | Dataset | Best Model | Paper | Code | Compare |
|---|---|---|---|---|---|
|
|||||
|
PACS
Most implemented papers
Learning to Diversify for Single Domain Generalization
busername/learning_to_diversify
•
•
ICCV 2021
Domain generalization (DG) aims to generalize a model trained on multiple source (i. e., training) domains to a distributionally different target (i. e., test) domain.
Adversarial Bayesian Augmentation for Single-Source Domain Generalization
Generalizing to unseen image domains is a challenging problem primarily due to the lack of diverse training data, inaccessible target data, and the large domain shift that may exist in many real-world settings.
Crafting Distribution Shifts for Validation and Training in Single Source Domain Generalization
The method that achieves the best performance on the augmented validation is selected from the proposed family.
