Pseudo Sample Synthesis Based on Task-Irrelevant Augmentation for Generalized Zero-Shot Learning

Generalized zero-shot learning try to recognizes images from seen and unseen classes. Recent methods construct two complementary visual representations, or directly disentangle the overall visual representations to synthesize pseudo samples to achieve it. But there are still two challenging problems: (1) Data-consuming. Constructing two complementary visual representations requires massive training data, which may cause poor per-formance when samples are limited. (2) Ignoring task-irrelevant features during synthesizing. The typical disentanglement-based methods synthesize pseudo samples without using learned task-irrelevant features. However, these features have better represen-tation transferability, which can contribute to more diversity syn-thesizing. In this paper, we propose the Pseudo Sample Synthesis based on Task-irrelevant Augmentation (PS<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>TA) that synthesizes by taking advantage of task-irrelevant features disentangled from global visual features, which has two key modules. (1) Anchor-pseudo-sample Synthesis Module, with which we synthesize class anchor pseudo samples that guarantee the conformity between pseudo samples and their corresponding semantics. (2) Diverse-pseudo-sample Synthesis Module. We disentangle task-irrelevant features from the global visual representations and synthesize diverse pseudo samples augmented by task-irrelevant features based on the anchor pseudo samples, which guarantees the diversity of pseudo samples. Extensive experiments on three benchmarks and different tasks demonstrate our method is competitive.