3 ms·
Around 2007, there was an image board type of site that was struggling to build a system that could reliably detect these sort of images. Their solution ended
by runlevel1 4y ago
Around 2007, there was an image board type of site that was struggling to build a system that could reliably detect these sort of images.
Their solution ended up being to run it through face detection first -- which was pretty reliable even back then.
It turns out, people who sent unsolicited lewd photos tended to do so without their face in frame. So if a face was detected, it was significantly less likely to be lewd.
Just an amusing anecdote. Wish I could find the blog post describing it.
- lucasfcosta 4y agoAn excellent example of cutting the Gordian Knot: https://en.wikipedia.org/wiki/Gordian_Knot https://en.wikipedia.org/wiki/Gordian_Knot
- z9znz 4y agoAre you making some kind of Freudian Slit?
- nerdponx 4y agoAlso a great example of effective feature engineering. The success of gradient boosting and deep learning lies largely in the ability of those models to "learn" sophisticated high-order features from raw data. Their particular success in working with images, audio, and video lie in their ability to construct more (and more sophisticated) features, beyond what a team of smart humans with domain knowledge could have constructed by hand given huge resources. In a sense, the only difference between model architectures is the feature space that they embed the raw data into. That said, sometimes you just need to give the model a good starting place, and sometimes you can obviate A sophisticated model largely or entirely, if you can come up with highly "explanatory" features on your own. To this day, I think the ability to come up with high-impact features is one of the differentiating factors of really good industry data scientist.