** by Justin Pinkney**

Original link /https://www.justinpinkney.com…

Download the dataset: V2

https://drive.google.com/file…

update

V2-28 images with poor quality (poorly aligned or no front) were removed.

V1 – The original release was used in the paper on resolution dependent GaN interpolation for controlled image synthesis between domains.

The data set

The Ukiyo-e face dataset contains 5,209 images of faces from Ukiyo-e photographs. The image pixels are 1024×1024, JPEG format, and have been aligned according to the procedure used for the FFHQ dataset. Above is a map of (almost) all the images in the dataset, plotted so that similar faces appear close together [2]. The image has been scaled down to 256×256 for easy display.

details

Images were downloaded from several museum websites, and I used Amazon Rekognition to try to detect faces and facial markers from each image. Rekognition works reasonably well on both tasks, but it’s clearly not perfect, with many faces missing and alignment errors in many of the images. At the same time, the resolution of many images is not very high, so in order to generate a usable data set with 1024×1024 resolution, I used the pre-trained ESRGAN [3] model trained on the MANGA109 data set to enlarge the images on demand. Although these images left some defects, the results were generally good.

Other data sets

Kaokore is another dataset of Ukiyo-e faces [4], which has more diversity and labels, but the image resolution is lower and the faces are not aligned.

License and Use

This data set is provided under the ‘Creative Commons Attribution – Share Like It 4.0 International License’.

If using a dataset, refer to as “Aligned ukiyo-e faces dataset, Justin Pinkney 2020” or a BibTeX entry:

  1. Pinkney, Justin N. M., Resolution Dependent GaN Interpolation for satellite-borne Image Synthesis Between Domains. ArXiv: 2010.05334 / Cs, Eess, 20 October 2020. http://arxiv.org/abs/2010.05334.
  2. To generate this image, I first extracted the features of CNN from each image using ResNet50 pretrained on ImageNet. Then UMAP is used to project these high-dimensional feature vectors to two-dimensional, and finally LAPJV algorithm is used to complete mesh division.
  3. Wang, Xintao, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Chen Change Loy, Yu Qiao, and Xiaoou Tang. Enhanced Super-Resolution Generative Adversarial Networks’. ArXiv:1809.00219 [Cs], 1 September 2018. http://arxiv.org/abs/1809.00219.
  4. Tian, Yingtao, Chikahiko Suzuki, Tarin Clanuwat, Mikel Bober-Irizar, Alex Lamb, and Asanobu Kitamoto. ‘Kaokore: A pre-modern Japanese Art Facial Expression Dataset ‘. ArXiv:2002.08595 [Cs, Stat], 20 February 2020. http://arxiv.org/abs/2002.08595.