Selective-generative feature representations for generalized zero-shot open-set classification by learning a tightly clustered space

Generating pseudo unseen samples is currently an effective approach for addressing the generalized zeroshot classification (GZSC) problem. However, in practical scenarios, the test set may contain open -set samples without semantic representations. Generalized zero -shot open -set classification (GZSOSC) extends GZSC tasks by simultaneously dealing with samples from seen, unseen, and open -set classes. In this paper, we propose a novel method to learn selective -generative feature representations (SGFR) to tackle the GZSOSC problem. Firstly, to handle the lack of unseen samples, we introduce a simple yet effective unseen feature generation method that leverages the seen-unseen relationships. Through an efficient alternating optimization strategy, we learn the seen-unseen relationships and the unseen visual centers. Secondly, to address the lack of open -set samples, we focus on learning a tightly clustered space for both seen and unseen classes. This enables effective open -set feature selection. We utilize the selected open -set samples to generate high -quality open -set features, thus enhancing the diversity of open -set samples. Extensive experiments are conducted to demonstrate the effectiveness of SGFR in handling the GZSOSC task.