FAR⁴: A Fast and Robust Classification Range Metric

This paper presents the Fast Approximation of the Raw Range-Recall Region (FAR4): a novel measure of the range at which a system can reliably classify a certain class of target. Specifically, we evaluate the metric using a dataset of radar images for drone classification. FAR<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">4</sup> calculates the area under the curve of a rank-based k-nearest neighbor regression of the range-recall curve. Because it is non-parametric, it doesn't re-quire any model assumptions. We show that it corresponds well with the range at which parametric models drop below 50% recall and that the metric is robust against pertur-bations in the hyperparameters and in the data. FAR<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">4</sup> can effectively be used to compare the performance of different sensors, classification algorithms, or neural networks. Since it natively computes the classification range for all possible probability score threshold values, users can effectively judge the classification ranges which correspond to the false alarm rate they are willing to tolerate.