Generalized Zero-Shot Learning With Multi-Channel Gaussian Mixture VAE
Jie Shao, Xiaorui Li · IEEE Signal Processing Letters · 2020
Generalized Zero-shot learning (GZSL) for object recognition is to solve the problem of recognizing samples in both seen and unseen classes, while only having seen classes in training. As labeling abundant examples for all kinds of classes in realistic scenarios is costly and impractical, GZSL has become a novel and hot field of research in recent years. Most previous methods tried to learn a fixed one-directional mapping, either from visual to semantic features, or from semantic to visual features. However, recently, cross mapping between visual and semantic features has achieved good results and many methods come up based on this idea. In this paper, we propose a novel model. It is Multi-channel Gaussian Mixture VAE(MCGM-VAE), which introduces Gaussian mixture model to our multi-modal VAE with multiple channels. These channels are of different weight coefficients following with channel-weight layers, so as to produce a Gaussian mixture distribution. Then the latent space could be generated from it. We evaluate our method on several benchmark databases, i.e. CUB, SUN, AWA1, AWA2, aPY and prove our approach outperforms state-of-the-art methods. Through the experimental data analysis, the impact of some hyperparameters on the experimental performance is further analyzed.