Contrastive visual feature filtering for generalized zero-shot learning

Generalized zero-shot learning aims to classify images from seen and unseen classes only by training with seen samples, which encounters the seen-unseen bias problem. Existing methods seek to solve the seen-unseen bias by synthesizing unseen samples. As only seen samples are involved during training, the synthetic unseen features tend to have the same distribution as the real visual features. Some redundant information in visual features is irrelevant to semantic description, so the synthetic unseen features generated based on these visual features also have redundant parts. In this paper, we propose a contrastive visual feature filtering framework (CVFF) for the generalized zero-shot learning task, eliminating redundant parts from both the real and the synthetic visual features. Specifically, a Feature Collaborative Filtering module (FCF) is proposed to filter out the relevant parts of visual features. To utilize the visual-semantic instance-level relationship, we introduce a visual semantic contrastive loss to optimize the model. Extensive experiments on multiple benchmarks for generalized zero-shot learning demonstrate that CVFF outperforms the state-of-the-art.