Zero-Shot Recognition through Image-Guided Semantic Classification
This addresses the problem of recognizing unseen classes in computer vision, but it is incremental as it builds on existing deep architectures.
The paper tackles zero-shot learning by proposing an embedding-based framework that inversely learns mapping between images and semantic classifiers, achieving state-of-the-art performance on standard benchmarks.
We present a new embedding-based framework for zero-shot learning (ZSL). Most embedding-based methods aim to learn the correspondence between an image classifier (visual representation) and its class prototype (semantic representation) for each class. Motivated by the binary relevance method for multi-label classification, we propose to inversely learn the mapping between an image and a semantic classifier. Given an input image, the proposed Image-Guided Semantic Classification (IGSC) method creates a label classifier, being applied to all label embeddings to determine whether a label belongs to the input image. Therefore, semantic classifiers are image-adaptive and are generated during inference. IGSC is conceptually simple and can be realized by a slight enhancement of an existing deep architecture for classification; yet it is effective and outperforms state-of-the-art embedding-based generalized ZSL approaches on standard benchmarks.