Zero-shot classification by transferring knowledge and preserving data structure

Abstract

In practical object recognition tasks, one often encounters a problem to recognize some unseen objects, which are some new categories without labeled images at training stage. For solving the challenging problem, zero-shot learning has been studied widely which can be seen as a special case of transfer learning. Thus, zero-shot learning deals with the problem of predicting labels of target images based on source images and their common semantic knowledge. Most existing zero-shot learning methods focus on how to project the images into the semantic space. However, the projection function learnt by source images and attributes has a shift for the prediction of target attributes. In this paper, we proposed a zero-shot classification method by transferring knowledge from source domain and preserving target data structure (TKDS). Particularly, we directly learn the target classification model by the semantic correlation with source classification model. Different from existing similarity based zero-shot learning methods, we also utilize the data properties of target data themselves. We simultaneously consider transferring knowledge from source domain to explicit the source images and utilizing the manifold structure of target data to rectify domain shift problem, thus, boosting the prediction performance. Extensive experiments on the widely used datasets show that our model outperforms significantly the state-of-the-art methods.