rstml
commented
5 years ago @TheDailySpank thanks for the feedback! Indeed, it's far from being ready for real-world applications. 95% accuracy is only against the test data set which is limited.
One way to tune classification is to tweak cut-off point, e.g. set it to 0.75 instead of 0.5. It's currently not supported in this Chrome extension, but we plan to add it in the future releases.
Our goal is to engage the community and iteratively improve model robustness. To do this, we open sourcing as much as we can. You can use our training scripts which you can apply to your dataset. See this repo for more details: https://github.com/purify-ai/geacc-models
We are preparing a new version trained on 100K images (twice of current model) and plan to release the dataset itself.
We also accept data donations. If you can share your dataset, we could include it in our next release. However, due to legal and technical limitations, we can only accept data in the form of publicly accessible URLs.
TheDailySpank
I've been working on a similar idea for a while and then stumbled upon this via another project I've been following.
While the method works (images deemed inappropriate get blurred and the tag works) the classification is way off and is less than the 95% claimed.
For example, an image of a woman in a vineyard, fully clothed, with a wine bottle to her mouth was highly scored while images I won't describe but were very obviously pornographic went right through.
There is also bias towards grids of images, probably because of the thumbnail sheets of adult videos causing similar image layouts to be classified incorrectly. e.g. This image scores .52 on the explicit scale. Don't worry, it's totally safe. I've seen this issue with Yahoo's OPEN_NSFW release.
I have a large dataset of manually corrected images if you'd be willing to give me info on how to train from it. Initial classification was done using Yahoo's model.