Novelty Detection in Thermal Video

A key limitation of deep neural networks (DNNs) is their tendency to predict high confidences when shown out-of-band input that differs from that trained on. Previous softmax probability based attempts to solve this problem have centered on synthetic test sets, drawn from significantly different distributions. We show the limitation of these methods when distributions lie on the same manifold and propose a density estimation based algorithm that increases the area under the receiver operating characteristic (AU-ROC) score from 0.640 to 0.802 on a real-world dataset.