Improving Network Architectures for Semantic Segmentation with Transformers

In this paper, we present SST, a simple and efficient framework for semantic segmentation that combines a Transformer-based encoder with a context-fusion decoder. Our framework uses a Transformer-based encoder to output features at multiple scales. We design a new decoder structure based on context fusion, which utilizes contextual information to fuse multiple feature maps at each pixel location and improve the semantic segmentation accuracy. Additionally, we achieve competitive results on several standard semantic segmentation benchmarks.