GLFormer: Efficient Global and Local Transformer with Lightweight Decoder for Semantic Segmentation

We propose an efficient and simple semantic segmentation model, termed as GLFromer. GLFomer has two contributions. Firstly, unlike the traditional hybrid Transformer-CNN structure, we propose the global and local context extractor, named as GL-Extractor. we combinate the self-attention and convolution in one layer, which not only significantly reduces the computation cost of model, but also captures the long context dependencies and local information at the same time. Second, we introduce the up sample lightweight decoder, where we handle the feature maps only by the concatenation and non-linear operation. The results of ablation and comparative studies can demonstrate the effectiveness of our structure.