22 May 2015 An evaluation of open set recognition for FLIR images
Author Affiliations +
Typical supervised classification algorithms label inputs according to what was learned in a training phase. Thus, test inputs that were not seen in training are always given incorrect labels. Open set recognition algorithms address this issue by accounting for inputs that are not present in training and providing the classifier with an option to reject" unknown samples. A number of such techniques have been developed in the literature, many of which are based on support vector machines (SVMs). One approach, the 1-vs-set machine, constructs a slab" in feature space using the SVM hyperplane. Inputs falling on one side of the slab or within the slab belong to a training class, while inputs falling on the far side of the slab are rejected. We note that rejection of unknown inputs can be achieved by thresholding class posterior probabilities. Another recently developed approach, the Probabilistic Open Set SVM (POS-SVM), empirically determines good probability thresholds. We apply the 1-vs-set machine, POS-SVM, and closed set SVMs to FLIR images taken from the Comanche SIG dataset. Vehicles in the dataset are divided into three general classes: wheeled, armored personnel carrier (APC), and tank. For each class, a coarse pose estimate (front, rear, left, right) is taken. In a closed set sense, we analyze these algorithms for prediction of vehicle class and pose. To test open set performance, one or more vehicle classes are held out from training. By considering closed and open set performance separately, we may closely analyze both inter-class discrimination and threshold effectiveness.
© (2015) COPYRIGHT Society of Photo-Optical Instrumentation Engineers (SPIE). Downloading of the abstract is permitted for personal use only.
Matthew Scherreik, Matthew Scherreik, Brian Rigling, Brian Rigling, "An evaluation of open set recognition for FLIR images", Proc. SPIE 9476, Automatic Target Recognition XXV, 94760N (22 May 2015); doi: 10.1117/12.2176585; https://doi.org/10.1117/12.2176585

Back to Top