Hacker News Reader

This is a great read. The OpenCV part hit a bit too close to home though: I was stuck for a minute trying to think how he managed to segment faces well enough to compute ratios (not just as a rectangle or a blob) given all of the possible conditions/perspectives of the photos.

View on HN

Clearly the right approach here would be crowdsourcing: submit all the pics using the hot-or-not API, keep only those with 7.5+ ratings.

you bet! i was automatically thinking about how did he manage to get sufficiently good resolution, how did he cope with lighting/background changes, and so on....

It's really the wrong approach. Supervised learning is the way to go. For example a paper by Kumar et al.[1] shows how to build an "attractive woman" classifier that is 83% accurate.

[1] http://homes.cs.washington.edu/~neeraj/publications/base/pap...

Presumably if they used 310 million faces instead of only 3.1 million it would be even more accurate, which is pretty impressive.

i'm confused - an .edu paper has been already mentioned while "Weird Science" hasn't been yet.

Reply on news.ycombinator.com