Cross-view classification by joint adversarial learning and class-specificity distribution

Highlights：

• We find that ℓ1,2-norm has an important role of characterizing class-specificity distribution in dimension space, which is not considered in the existing methods. Therefore, we apply it to learn latent representations which well characterize class structure.

• We integrate adversarial learning and class-specificity into a unified optimization framework. Thus, the learned view-consistent representations not only well encode discriminant information but also well characterize class structure in dimension space.

• Extensive experiments on several datasets are conducted to demonstrate the superiority of our method over state-of-the-art methods, and the ablation experiments prove the effectiveness of our contributions.

摘要

•A novel deep adversarial network is proposed for cross-view classification. In order to solve the scale-sensitive problem in existing methods, adversarial learning is used to get a robust view-consistent representation.•We find that ℓ1,2-norm has an important role of characterizing class-specificity distribution in dimension space, which is not considered in the existing methods. Therefore, we apply it to learn latent representations which well characterize class structure.•We integrate adversarial learning and class-specificity into a unified optimization framework. Thus, the learned view-consistent representations not only well encode discriminant information but also well characterize class structure in dimension space.•Extensive experiments on several datasets are conducted to demonstrate the superiority of our method over state-of-the-art methods, and the ablation experiments prove the effectiveness of our contributions.