janvanlooy commited on
Commit
5d22c29
1 Parent(s): f5edb1a

Various changes

Browse files
Files changed (1) hide show
  1. README.md +9 -8
README.md CHANGED
@@ -119,17 +119,18 @@ python extract_images.py --parquet_file <Path to the Parquet file or folder cont
119
  If you want to contribute to the dataset, the best way is to help us develop pipeline components for further processing. Components
120
  we are currently looking to add are the following ([GitHub issues](https://github.com/ml6team/fondant/issues?q=is%3Aissue+is%3Aopen+label%3A%22Component+Contribution%22)):
121
  - Image-based deduplication
122
- - Visual quality / aesthetic quality estimation
123
  - Automatic captioning
124
- - Not safe for work (NSFW) content detection
125
  - Watermark detection
 
 
126
  - Face detection
127
  - Personal Identifiable Information (PII) detection
128
  - Text detection
129
  - AI generated image detection
130
- - CLIP embedding generation
131
  - Image-text CLIP similarity
132
- - ...
 
133
  We are also looking for core framework contributors and users who are willing to give feedback on usability and suggest potential improvements
134
 
135
 
@@ -165,11 +166,11 @@ fondant-cc-25m is built from CommonCrawl dumps. These dumps are constructed from
165
  Permissive licenses have minimal restrictions on how the image can be copied, modified, and redistributed.
166
  The full list of licenses can be found [here](https://creativecommons.org/about/cclicenses/).
167
  We examined HTML tags of the webpages for the presence of Creative Commons license URLs. A webpage was marked permissive only when a license URL was found in
168
- its footer, aside or sidebar. This was the case only in around 0.164% of a 100k sample set which suggests that image generation models trained on a random sample from
169
- the public internet may be trained on up to 99.836% images that may be copyrighted.
170
 
171
- Subsequently, all the image URLs present on the web page were collected together with the license information.
172
- A manual test of 1032 randomly sampled images showed an accuracy of 96.32% in which case the image was actually released under a Creative Commons license.
173
  False positives could be due to parsing errors but also incorrect attributions: images indicated by the publisher to be CC which are not.
174
  More information on our approach can be found in [this blogpost](https://blog.ml6.eu/ai-image-generation-without-copyright-infringement-a9901b64541c).
175
 
 
119
  If you want to contribute to the dataset, the best way is to help us develop pipeline components for further processing. Components
120
  we are currently looking to add are the following ([GitHub issues](https://github.com/ml6team/fondant/issues?q=is%3Aissue+is%3Aopen+label%3A%22Component+Contribution%22)):
121
  - Image-based deduplication
 
122
  - Automatic captioning
123
+ - Visual quality / aesthetic quality estimation
124
  - Watermark detection
125
+ - Not safe for work (NSFW) content detection
126
+ - CLIP embedding generation
127
  - Face detection
128
  - Personal Identifiable Information (PII) detection
129
  - Text detection
130
  - AI generated image detection
 
131
  - Image-text CLIP similarity
132
+ - Any components that you propose to develop\
133
+
134
  We are also looking for core framework contributors and users who are willing to give feedback on usability and suggest potential improvements
135
 
136
 
 
166
  Permissive licenses have minimal restrictions on how the image can be copied, modified, and redistributed.
167
  The full list of licenses can be found [here](https://creativecommons.org/about/cclicenses/).
168
  We examined HTML tags of the webpages for the presence of Creative Commons license URLs. A webpage was marked permissive only when a license URL was found in
169
+ its footer, aside or sidebar. This was the case only in around 0.164% of a 100k random sample from Common Crawl. This suggests that image generation models
170
+ trained on a random sample from the public internet may be trained on up to 99.836% copyrighted images.
171
 
172
+ Subsequently, all the image URLs present on the web page were collected together with the license information. A manual check of a random
173
+ sample of 1032 images showed that 96.32% were attributed the correct license whil 3.68% were not.
174
  False positives could be due to parsing errors but also incorrect attributions: images indicated by the publisher to be CC which are not.
175
  More information on our approach can be found in [this blogpost](https://blog.ml6.eu/ai-image-generation-without-copyright-infringement-a9901b64541c).
176