Transductive zero-shot learning via knowledge graph and graph convolutional networks.

Zero-shot learning methods are used to recognize objects of unseen categories. By transferring knowledge from the seen classes to describe the unseen classes, deep learning models can recognize unseen categories. However, relying solely on a small labeled seen dataset and the limited semantic relationships will lead to a significant domain shift, hindering the classification performance. To tackle this problem, we propose a transductive zero-shot learning method, based on Knowledge Graph and Graph Convolutional Network. We firstly learn a knowledge graph, where each node represents a category encoded by its semantic embedding. With a shallow graph convolutional network having a small number of layers, we learn the classifier for each category, supervised by the visual classifiers of the seen categories. During testing, a clustering strategy, the Double Filter Module with Hungarian algorithm, is applied to the unseen samples, and then, the learned classifiers are used to predict their categories. Pseudo annotations are given to the samples that are more accurately classified. In the transductive setting, the unseen categories with higher classification accuracy are assigned pseudo annotations, and can be associated with the seen categories to progressively update model parameters. We validate the proposed model on three data sets, and our model outperforms other state-of-the-art methods, achieving 47.36% accuracy on AWA2, 30.69% on ImageNet50, and 18.87 on ImageNet100, with a 4-10% improvement over existing methods.