I wonder how a visual transformer network would handle a problem like this. I imagine that it’s use of attention mechanisms would make it easier for it to spot porn
@Ganondox
At the bottom you see Equal Error Rate 9.48% (point when false positive rate = false negative rate). Indeed at this point the filter would not be the most accurate thing ever. However without more information we can’t say how well it would work if you set it very sensitive to detecting explicit images (we need ROC curve for that).
Ah yes, my bad! I thought I was commenting on the [suggestive] version of the image ^^.
Yes, it is a non-explicit image which got high score for explicit. tag: "explicit" means the saliency map is computed for tag [explicit]. ittedly it is rather suggestive image, which is probably why it got high score. This image was not in the training set.
That’s not what I meant, although I may have misunderstood. >>1748971, here on Derpibooru, is not tagged “explicit.” In the image above, you say ‘tag: “explicit”’ underneath it. Do you mean that this is a non-explicit image which the program improperly thought was explicit? I just want to make sure a teaching image didn’t get miscategorized.
@dashbug
Much like #9793 said, that model is not even aware of clop images and hence gives nonsense with such images.
I run this model here on the images you posted, and got 99% and 61% “explicit-percentage”.
(Sidenote: Ponies could make more interesting nsfw-detection dataset, because ponies come in different colors and shapes. Last time I checked, isitporn.com seems to focus on naked-skin color. We would rather want it to find explicit bits, much like this model did with the cherry-picked examples)
Could be pretty useful to… make more SFW filtering content…
My, 2 years already, I now lean towards data visualization rather than to train the model lately.
At the bottom you see Equal Error Rate 9.48% (point when false positive rate = false negative rate). Indeed at this point the filter would not be the most accurate thing ever. However without more information we can’t say how well it would work if you set it very sensitive to detecting explicit images (we need ROC curve for that).
This is hilarious!
Ah yes, my bad! I thought I was commenting on the [suggestive] version of the image ^^.
Yes, it is a non-explicit image which got high score for explicit.
tag: "explicit"means the saliency map is computed for tag [explicit]. ittedly it is rather suggestive image, which is probably why it got high score. This image was not in the training set.That’s not what I meant, although I may have misunderstood. >>1748971, here on Derpibooru, is not tagged “explicit.” In the image above, you say ‘tag: “explicit”’ underneath it. Do you mean that this is a non-explicit image which the program improperly thought was explicit? I just want to make sure a teaching image didn’t get miscategorized.
Much like #9793 said, that model is not even aware of clop images and hence gives nonsense with such images.
I run this model here on the images you posted, and got 99% and 61% “explicit-percentage”.
(Sidenote: Ponies could make more interesting nsfw-detection dataset, because ponies come in different colors and shapes. Last time I checked, isitporn.com seems to focus on naked-skin color. We would rather want it to find explicit bits, much like this model did with the cherry-picked examples)
That’s because it’s trained to detect IRL porn, not clop.
I have such the best excuse if anyone finds my stash now
Examples:
http://isitporn.com/show/9a5c12fa91a7400d9d5ecaa8a5cdbe2d
http://isitporn.com/show/d9ae26db4104466f89eb64d54f5312b7
These explicit images are believed to be safe.
… But my cat images (even with only visible face) are almost nsfw
Edited
Indeed it isn’t. I first aimed to tag this [safe] (hence the censor), but ultimately sticked with [suggestive] but kept the censor-block ^^’
This is almost as impressive as that AI that writes commercials for Olive Garden.