No abstract

Landmine detection methods face the challenging task of discriminating a wide variety of targets in a diverse array of environmental conditions. As a result, successful approaches often utilize multiple sensing modalities and algorithms in order to overcome the limitations of any one particular approach. The development of complementary approaches is often motivated by a reductionist view, which reduces a detector's overall limitations to its limitations within particular contexts. However, when a detector is evaluated on a diverse data set, a new source of error arises that is not reflected in its performance within the contexts in isolation, which we refer to as mis-calibration error. Mis-calibration occurs when the true likelihood of target presence conditioned on a detector's output significantly varies across different contexts. In this work, we demonstrate the effect of context-dependent mis-calibration first on a hypothetical detector and then on four different landmine detection methods. To eliminate error due to mis-calibration, we propose a strategy called context-dependent score calibration (CDSC), which uses a monotonic calibration approach that maximizes the AUC (area under the ROC curve) of the calibrated output. We then show the gain in observed performance achieved on a diverse dataset when these four landmine detection algorithms are properly calibrated across the different contexts.