narugo's picture
Update README.md
4746a00
---
license: mit
task_categories:
- zero-shot-classification
tags:
- art
size_categories:
- 100K<n<1M
---
# character_similarity
This is a dataset used for training models to determine whether two anime images (containing only one person) depict the same character. The dataset includes the following versions:
| Version | Filename | Characters | Images | Information |
|:---------:|:-----------------------:|:----------:|:------:|---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
| v0 | images_v0.tar.xz | 2059 | 162116 | Crawled from [zerochan.net](https://www.zerochan.net/), includes images of Arknights, Fate/Grand Order, Genshin Impact, Girls' Frontline, and Azur Lane, as well as over 1500 other game or anime characters. The images are all small preview thumbnails. |
| v0_tiny | images_tiny_v0.tar.xz | 514 | 10036 | The dataset `v0` has been simplified by randomly selecting 1/4 of the anime characters and 1/4 of the images for each character. It is recommended to use this dataset for model training and validation. |
| v0_xtiny | images_xtiny_v0.tar.xz | 100 | 1814 | `v0_xtiny` dataset is a further simplified version of `v0_tiny`, with only 100 characters retained. This dataset is only **suitable for model validation and experimentation**, and is not recommended for formal training. |
| v1 | images_v1.tar.xz | 4001 | 292800 | The `v1` dataset is similar to `v0`, crawled from zerochan, but it contains a richer selection of characters and more diverse images of the same character. Each character includes no less than 3 and no more than 200 images. |
| v1_pruned | images_pruned_v1.tar.xz | 3982 | 241483 | Pruned from `v1` dataset, deleted monochrome and non-solo pictures, and also removed pictures where the character accounted for less than 40% of the entire vision area. |
| v2 | images_v2.tar.gz | 37886 | 836217 | Images containing only the faces of all characters from [zerochan.net](https://zerochan.net) have been scraped, and the image size is larger than that of the v1 and v0 datasets. However, please note that the v2 dataset may include multi-level nested image paths, such as `a/b/c/xx.jpg` and `a/b/yy.jpg`. In such cases, `xx.jpg` and `yy.jpg` should not be considered the same character. **Two characters from different images are considered the same character only if they are located in exactly the same path.** |