Open Set Recognition with Generated Data

Lawrence Neal, Matthew Olson, Weng-Keen Wong, Xiaoli Fern, & Fuxin Li

Abstract: In open set recognition tasks, a classifier must label instances of known classes while detecting unknown classes not encountered during training. We propose a framework in which a classifier jointly learns to classify instances of known classes, and to detect unknown classes as an additional "open set" class. Training examples for this open set class are synthesized by a generative adversarial network, itself trained only on the known classes. Augmenting the dataset with synthesized open set examples improves upon standard techniques like confidence thresholding.

Keywords: Open Set Recognition, Generative Adversarial Network, Dataset Augmentation

TL;DR: For learning the open set, some GAN strategies work better than others.