[ebarto]

Well, thanks. So:

1) We are using rapid Object Detection by training Haar classifiers of different bikini shapes (top, bottom, strapless bikinis etc). We used opencv_haartraining to get a few XMLs we use for testing each image.

2) The true positive and false positive rates vary depending on the test we are performing (top, bottom...) so we have a formula based on the combination of these and the amount of hits found in each picture by each classifier, where we assign the picture a probability of it containing a bikini (based on a controlled sample of 3000 pictures). As of now we are looking at user feedback to make sure our threshold is where it needs to be so that we provide just the right amount of true/false positives in our results.