Joint discriminative attributes and similarity embeddings modeling for zero-shot recognition

Abstract

Zero-shot recognition (ZSR) is an appealing technique when addressing novel categories without any training data. However, existing methods face two critical limitations: 1) most existing works directly exploit user-defined attributes to transfer knowledge across classes, yet the importance to learn discriminative attributes is neglected; 2) previous studies typically consider the semantic similarity embedding, while there are limited research efforts on exploring the visual similarity embedding. In this paper, we propose a novel approach for ZSR to address the above issues. Specifically, we develop a simple yet effective strategy to learn discriminative attributes via regressing from labels, which yields superior performance in the semantic similarity space. Meanwhile, to our best knowledge, we are among the first to consider the visual similarity embedding, which fully explores the discriminative information of visual features to benefit the final performance. Moreover, we combine the visual and semantic similarity embeddings, which enriches the information in the joint similarity space and greatly reduces the gap between visual features and semantic attributes. Extensive experiments on five benchmarks show that our method achieves state-of-the-art results on both conventional and generalized zero-shot recognition tasks.