Visual representation learning is the key of solving various vision problems. Relying on the seminal grid structure priors, convolutional neural networks (CNNs) have been the de facto standard architectures of most deep vision models. For instance, classical semantic segmentation methods often adopt a fully-convolutional network (FCN) with an encoder-decoder architecture. The encoder progressively reduces the spatial resolution and learns more abstract visual concepts with larger receptive fields. Since context modeling is critical for segmentation, the latest efforts have been focused on increasing the receptive field, through either dilated (i.e., atrous) convolutions or inserting attention modules. However, the FCN-based architecture remains unchanged. In this paper, we aim to provide an alternative perspective by treating visual representation learning generally as a sequence-to-sequence prediction task. Specifically, we deploy a pure Transformer to encode an image as a sequence of patches, without local convolution and resolution reduction. With the global context modeled in every layer of the Transformer, stronger visual representation can be learned for better tackling vision tasks. In particular, our segmentation model, termed as SEgmentation TRansformer (SETR), excels on ADE20K (50.28% mIoU, the first position in the test leaderboard on the day of submission), Pascal Context (55.83% mIoU) and reaches competitive results on Cityscapes. Further, we formulate a family of Hierarchical Local-Global (HLG) Transformers characterized by local attention within windows and global-attention across windows in a hierarchical and pyramidal architecture. Extensive experiments show that our method achieves appealing performance on a variety of visual recognition tasks (e.g., image classification, object detection and instance segmentation and semantic segmentation).
* Extended version of CVPR 2021 paper arXiv:2012.15840