Zero-shot semantic segmentation using relation network

Abstract

Zero-shot learning (ZSL) is widely studied in recent years to solve the problem of lacking annotations. Currently, most studies on ZSL are for image classification and object detection. But zero-shot semantic segmentation, pixel level classification, is still at its early stage. Therefore, this thesis proposes to extend a zero-shot image classification model, Relation Network (RN), to semantic segmentation tasks. This thesis modifies the structure of RN based on other state-of-the-arts semantic segmentation models (i.e. U-Net and DeepLab) and utilizes word embeddings from Caltech-UCSD Birds 200-2011 attributes and natural language processing models (i.e. word2vec and fastText). Because meta-learning is limited to binary tasks, this thesis proposes to join multiple binary semantic segmentation pipelines for multi-class semantic segmentation. It is proved by experiments that RN could improve accuracy of U-Net with the help of semantic side information on binary semantic segmentation and it could also be applied on multi-class semantic segmentation with simpler structure than the baseline model, SPNet, but higher accuracy under ZSL setting. However, the capability of RN under generalized zero-shot learning (GZSL) setting still needs improvement. This thesis also studies on how different word embeddings, network structures and data affect RN and what could be done to improve its results.