Self-supervised vision transformers for semantic segmentation

Abstract

Semantic segmentation is a fundamental task in computer vision and it is a building block of many other vision applications. Nevertheless, semantic segmentation annotations are extremely expensive to collect, so using pre-training to alleviate the need for a large number of labeled samples is appealing. Recently, self-supervised learning (SSL) has shown effectiveness in extracting strong representations and has been widely applied to a variety of downstream tasks. However, most works perform sub-optimally in semantic segmentation because they ignore the specific properties of segmentation: (i) the need of pixel level fine-grained understanding; (ii) with the assistance of global context understanding; (iii) both of the above achieve with the dense self-supervisory signal. Based on these key factors, we introduce a systematic self-supervised pre-training framework for semantic segmentation, which consists of a hierarchical encoder–decoder architecture MEVT for generating high-resolution features with global contextual information propagation and a self-supervised training strategy for learning fine-grained semantic features. In our study, our framework shows competitive performance compared with other main self-supervised pre-training methods for semantic segmentation on COCO-Stuff, ADE20K, PASCAL VOC, and Cityscapes datasets. e.g., MEVT achieves the advantage in linear probing by +1.3 mIoU on PASCAL VOC.