Unpaired Apple and Orange Images for CycleGAN

A collection of unpaired images of apples and oranges, originally part of the CycleGAN benchmark suite. It was created by Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros for research on unpaired image-to-image translation. The dataset is hosted by the 'huggan' organization and was last updated in April 2022.

Use Cases

• Train a CycleGAN model to translate images from the apple domain to the orange domain and vice versa.
• Benchmark unpaired image-to-image translation algorithms using the apple and orange categories.
• Analyze the visual features distinguishing apple images from orange images for domain adaptation research.

Strengths

• Dataset is associated with the seminal 2017 CycleGAN paper (arXiv:1703.10593).
• Part of a well-known benchmark suite for image-to-image translation tasks.
• Hosted on a platform that supports libraries like Polars, Pandas, and Datasets for accessibility.

Limitations

• The exact number of images, file formats, and image resolution are not specified in the provided input.
• Dataset size is categorized as '1K<n<10K', indicating it contains between 1,000 and 10,000 items, but the precise count is unknown.
• Limited to only two object categories (apples and oranges), which restricts broader multi-class applications.

Provenance

Source

Originally from the CycleGAN datasets hosted by researchers at UC Berkeley.

Collection Method

Curated collection of images for unpaired image-to-image translation research.

Freshness

Last updated on 2022-04-12.

Geography

Region tag suggests a US focus, but specific coverage is unknown.