An Iterative Co-Training Transductive Framework for Zero Shot Learning

In zero-shot learning (ZSL) community, it is generally recognized that transductive performs better than inductive one as the unseen-class samples are also used in its training stage. How to generate pseudo labels for and how use such usually noisy two critical issues learning. this work, we introduce an iterative co-training framework which contains different base ZSL models exchanging module. At each iteration, co-trained separately predict samples, module exchanges predicted labels, then exchanged pseudo-labeled added into sets next iteration. By such, our can gradually boost performance by fully exploiting potential complementarity of models' classification capabilities. addition, applied generalized (GZSL), a semantic-guided OOD detector proposed pick out most likely before class-level alleviate bias problem GZSL. Extensive experiments on three benchmarks show methods could significantly outperform about 31 state-of-the-art ones.