Hi Gvelius,

Thanks for your interest!

I am the original author of the library, which was published anonymously to accommodate the peer review process of the attached paper.

My apologies for the late reply. Apparently I no longer receive email notifications from this repo.

This is an obvious feature that I will look at implementing. I need to update the pypi package anyway, as apparently there are no builds for python > 3.6.

For the time being I would suggest the following work around, which takes into account the fact that FMRs and FNMRs are basically indices on the genuines and (inversed) impostor score lists, if these lists are sorted. Thus, EERs are as well.

import math

import numpy as np


def eer_to_threshold(eer, impostors, genuines):

    # as the EER is interpolated, we need to reverse the "bounding box" indices,
    # making sure we don't try out of bound reads from the list
    g_lower_bound = min(math.floor(eer * len(genuines)), len(genuines) - 1)
    g_upper_bound = min(math.ceil(eer * len(genuines)), len(genuines) - 1)

    i_lower_bound = min(math.floor((1 - eer) * len(impostors)), len(impostors) - 1)
    i_upper_bound = min(math.ceil((1 - eer) * len(impostors)), len(impostors) - 1)

    score_glb = genuines[g_lower_bound]
    score_gub = genuines[g_upper_bound]

    score_ilb = impostors[i_lower_bound]
    score_iub = impostors[i_upper_bound]

    # we need to determine the smallest range within which the threshold can lie. We take the max of the lower bounds of the scores
    # the mins of the upper bounds of the scores.
    lb = max(score_glb, score_ilb)
    ub = min(score_gub, score_iub)
    # Interpolate between the bounds to determine the treshold.
    return (lb + ub) / 2


genuines = np.linspace(
    0, 1, 100
)  # insert your sorted list of genuine (similarity) scores
impostors = np.linspace(
    -0.5, 0.5, 100
)  # insert your sorted list of impostor (similarity) scores
eer = 0.2527  # calculated with e.g. `feer`
# eer = 1

t = eer_to_threshold(eer, impostors, genuines)
print("threshold = ", t)

I will keep this issue open until I implement the new functionality.

is it possible to retrieve the threshold (i.e. similarity/dissimilarity) at the EER? #1

Description

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions