Zero-shot learning (ZSL) is a challenging task aiming at recognizing novelclasses without any training instances. In this paper we present a simple buthigh-performance ZSL approach by generating pseudo feature representations(GPFR). Given the dataset of seen classes and side information of unseenclasses (e.g. attributes), we synthesize feature-level pseudo representationsfor novel concepts, which allows us access to the formulation of unseen classpredictor. Firstly we design a Joint Attribute Feature Extractor (JAFE) toacquire understandings about attributes, then construct a cognitive repositoryof attributes filtered by confidence margins, and finally generate pseudofeature representations using a probability based sampling strategy tofacilitate subsequent training process of class predictor. We demonstrate theeffectiveness in ZSL settings and the extensibility in supervised recognitionscenario of our method on a synthetic colored MNIST dataset (C-MNIST). Forseveral popular ZSL benchmark datasets, our approach also shows compellingresults on zero-shot recognition task, especially leading to tremendousimprovement to state-of-the-art mAP on zero-shot retrieval task.
展开▼
