The hinge transform for writer identification / Equal-error rates example

The curve below (and a zoomed-in version) show where the equal-error rate is, i.e., the crossing point between the cumulative distributions of wrongly accepting (Pdiff) and wrongly rejecting (Psame) a pair of documents as coming from the same writer. The curves intersect at an EER of 3% for this implementation of the hinge, which was convolutional, over edges in the image, using 31 bins and legs of 7 pixels. Of the full 31x31 angle-combination matrix, a subset of 387 angle combinations is reached with legs of this size and are sensible (some redundant combinations are left out). Results are on the Firemaker data set (250 writers, 2 paragraphs/writer).