Hierarchical-Dynamic Embedding for Zero-Shot Object Recognition

Zero-shot object recognition is aiming to attach unseen category labels to images which are out of the training set. The key challenge in Zero-shot learning is building the map between visual domain and semantic domain. However, previous Visual-Semantic Embedding ignores the essential difference between the vectors of category names and the vectors of the entities. Hybrid model, moreover, computes the middle vector with a fixed size candidate set which limits the generalization on different images. So we propose a novel framework named Hierarchical-Dynamic Embedding. First, Hierarchical Network Embedding (HNE) takes advantage of the internal hierarchical taxonomy of the category names. We then provide Dynamic Hybrid Model (DHM) to map unseen images from visual vectors to entity vectors. Furthermore, we conduct the experiments on 1,000 seen categories and 1,548 unseen categories to show the state-of-the-art performance of our proposed framework.