task_categories:
- image-segmentation
size_categories:
- 10K<n<100K
This dataset consists of 26,000 anime style images, half of which are foreground characters or objects, and the other half of which are backgrounds. It is intended for training segmentation or matting models where the foreground subject can be extracted from the background. The foundation of this dataset is based upon https://huggingface.co/datasets/skytnt/anime-segmentation
I found the overall quality of that dataset did not meet my needs, so I did a lot of automated and manual inspection of the images, resulting in removing more than half of them, and then adding many more new images.
For the foreground images, I have removed ones containing nudity or extreme lewdness. I also carefully examined the images to remove ones containing the following issues: stray pixels in the image or the alpha channel, images that are cut off at the edge of the frame, semi-transparent areas, fuzzy/blurry areas in the alpha channel, drop shadows, text and other unrelated items appear in the image, partial backgrounds are behind the characters. The foreground images are mostly taken from booru image sites, but I have also added some sprites from games and visual novels, as well as images from various "transparent png" archives. I also tried to bring in more images of male characters and non-human creatures.
For the background images, I wanted to ensure that there was a stronger representation of backgrounds from actual anime videos. The original anime-segmentation dataset contained many abstract backgrounds and patterns, and most of the images were from anime-style illustrations rather than anime video. While some of those images have been preserved, I removed a large portion of them, and replaced them with backgrounds obtained from a variety of other sources, such as torrents and twitter, as well as manually capturing images from numerous anime videos myself. I also added a number of solid color background images, but these are located at the very end of the dataset and can be easily removed if you wish.
Across the entire dataset, I have also made efforts to remove duplicate and similar images.