On Implicit Attribute Localization for Generalized Zero-Shot Learning

Zero-shot learning (ZSL) aims to discriminate images from unseen classes by exploiting relations seen via their attribute-based descriptions. Since attributes are often related specific parts of objects, many recent works focus on discovering discriminative regions. However, these methods usually require additional complex part detection modules or attention mechanisms. In this paper, 1) we show that common ZSL backbones (without explicit nor detection) can implicitly localize attributes, yet property is not exploited. 2) Exploiting it, then propose SELAR, a simple method further encourages attribute localization, surprisingly achieving very competitive generalized (GZSL) performance when compared with more state-of-the-art methods. Our findings provide useful insight for designing future GZSL methods, and SELAR provides an easy implement strong baseline.