Episode-based Training Strategy for Zero-Shot Semantic Segmentation

We introduce episode-based training into zero-shot semantic segmentation (ZS3) for the first time. In particular, the model is trained on a set of simulated ZS3 tasks. The model gains the ability to predict simulated unseen classes over multiple episodes, which generalizes well to true unseen classes after training on multiple episodes. On the basis of this training framework, we propose a visual semantic alignment network named VSAN as our basic model, which mainly includes a feature extractor and semantic projection network. This base model constrains the visual-semantic distribution of the same class by a distance measure between visual features and semantic prototypes. In the inference phase, the trained projection network can generate corresponding semantic prototypes for all classes, and predict the segmentation results of the entire image by measuring the distance between visual features and semantic prototypes. The base model VSAN is called EB-VSAN after using the episode-based training strategy. Our model is a discriminative model, as opposed to generative methods, our model does not need retraining the classifier when a new class emerges, while avoiding multi-stage training. Our extensive ZS3 experiments on the benchmark dataset show that the EB-VSAN model outperforms current state-of-the-art methods, specifically, our hIoU metric outperforms state-of-the-art methods by an average of 1.7% on PASCAL VOC, while Average 2.5% improvement on PASCAL Context.