Zero-Shot Semantic Segmentation via Variational Mapping

We have witnessed the explosive success of deep neural networks (DNNs). However, DNNs typically assume a large amount of training data, and this is not always available in practical scenarios. In this paper, we present zero-shot semantic segmentation, where a model that has never seen the target class during training. For this purpose, we propose variational mapping, which facilitates effective learning by mapping the class label embedding vectors from the semantic space to the visual space. Experimental results using Pascal VOC 2012 show that our proposed method can achieve a mean intersection over union (mIoU) of 42.2, and we believe that this can serve as a baseline for similar research in the future.