Solving the Defect in Application of Compact Abating Probability to Convolutional Neural Network Based Open Set Recognition

Abstract

Close set is a hypothesis utilized by the majority of machine-learning-based (ML-based) recognition algorithms, assuming all testing classes are known at training time. In real world, the more practical model is Open Set Recognition (OSR), which allows the presence of unknown classes at testing time, but requires the rejection ability of the model. The compact abating probability (CAP) model, which assumes the probability of class membership decreases in value (abates) as points move from known data toward open space, is first raised in traditional ML-based OSR method and soon become the basis of majority of later developed works. Most of convolutional-neural-network-based (CNN-based) OSR methods also adopted this model as their basis. During our exploration, however, we find that the application of CAP model to the CNN-based OSR method is restricted by the difference of its feature space from that of ML-based method. To the best of our knowledge, we are the first group who find this gap. To fill this gap, we propose a method called OpenSoftMax to transform the CNN-based methods' features by the process of SoftMax. In order to investigate performance, we further implement quantitative comparison between our OpenSoftMax method and the well-known CNN-based method OpenMax on caltech256 datasets. Extensive experiments have been conducted to verify the effectiveness and efficiency of our proposals.