|
--- |
|
task_categories: |
|
- image-classification |
|
size_categories: |
|
- 10K<n<100K |
|
--- |
|
# Danbooru SFW 512 General Filter Dataset V1 |
|
This dataset is meant to be used for training a simple binary classifier that can filter the |
|
Danbooru SFW 2021 dataset. There are two classes: "accepted" and "rejected", with "accepted" |
|
representing samples that should pass through the filter and "rejected" representing samples |
|
that should not. |
|
|
|
Generallly, to be accepted a sample should meet the following criteria: |
|
|
|
1. No horizontal/vertical bars, or bars that can be removed easily by filling in the background |
|
color. This precludes any samples with bars which cut through any part of the image that is not |
|
a solid color block (i.e. patterns, detailed backgrounds, and foreground elements). |
|
|
|
2. Not overtly sexual. The SFW subset of Danbooru 2021 which this dataset is based on has |
|
already removed almost all nudity, this criterion simply removes some of the borderline NSFW images. |
|
|
|
3. Somewhat aesthetic. |
|
|
|
Please note that dataset samples were selected manually, so there was somewhat of a tradeoff between |
|
speed and accuracy. Despite this, I've managed to train a binary classifier up to 85% test |
|
accuracy on this dataset. I've also observed pretty good performance when using that classifier on |
|
entirely new portions of the base dataset. |
|
|
|
## Original Dataset Citation |
|
```bibtex |
|
@misc{danbooru2021, |
|
author={Anonymous and Danbooru community and Gwern Branwen}, |
|
title={Danbooru2021: A Large-Scale Crowdsourced and Tagged Anime Illustration Dataset}, |
|
howpublished={\url{https://gwern.net/danbooru2021}}, |
|
url={https://gwern.net/danbooru2021}, |
|
type={dataset}, |
|
year={2022}, |
|
month={January}, |
|
timestamp={2022-01-21}, |
|
note={Accessed: 2023-12-06} |
|
} |
|
``` |