About

Stylized Meta-Album (SMA) is a new meta-dataset, consisting of 24 datasets(12 content datasets and 12 stylized datasets), designed to enhance studies on out-of-distribution generalization and related topics. SMA is constructed using style transfer techniques from 12 subject classification datasets (content datasets), SMA provides a diverse and extensive set of 4800 groups combining various subjects (objects, plants, animals, human actions, textures) with multiple styles. This scale facilitates rigorous and expansive research that traditional datasets, which offer limited group variability and/or class diversity, cannot support. SMA is a continuously growing meta-dataset. See our datasets in Datasets Section.

We repurposed datasets that were generously made available by Meta-Album , see credits page. All datasets are free for use for academic purposes, provided that proper credits are given. For your convenience, you may cite our paper, which references all original creators.

We have summarized the data generation in the flow diagram below:

License

Stylized Meta-Album is released under a CC BY-NC 4.0 license permitting non-commercial use for research purposes, provided that you cite us. Additionally, redistributed datasets have their own license, see the credits page. All resources made available through this website provided “as is”. The curators of Stylized Meta-Album (and their home institutions and their sponsors) who have worked on its preparation, this website, the code provided to read, process data, and run baseline methods, make no warranties concerning the licensed material, including fitness for any purpose, non-infringement absence of defects or errors, accuracy, and they decline any liability for losses or other possible consequences that may arise by using such material.
This briefly summarizes the terms of the license CC BY-NC 4.0 and the disclaimer (that the license includes).

Recommended use

The recommended use of Stylized Meta-Album is to conduct fundamental research on machine learning algorithms and conduct benchmarks, particularly in: group fairness, few-shot learning, meta-learning, continual learning, transfer learning, and image classification.

Code

We provide code in our GitHub Repository for

1. Data processing
2. Data formatting
3. Quality control
4. Stylized Meta-Album use cases

Visit our GitHub repository for more details.

Stylized Meta-Album GitHub Repository

Datasets

We list in the tables below the data statistics of the SMA datasets. We chose 12 datasets from Meta-Album (content datasets). We scrapped a style dataset from the internet and we stylized the content datasets with the style dataset to get 12 stylized datasets. Each instance in these datasets is an image of 256x256 pixel images. Both Content and Stylized datasets are available in 2 versions:

1. Meta-Album Extended: all classes and all images per class
2. Stylized Meta-Album Mini: same as Stylized Meta-Album Extended, but we randomly sampled only 40 examples for each class (hence the datasets are class-balanced).

Gallery

Some sample images from Stylized Meta-Album datasets. Row 1: content images. Row 2: content images stylized by style images. Each image represents a sample from one dataset

Content Images: original images from Meta-Album datasets
Stylized Images: content images stylized by style images