With ViT as a backbone showing great promise, various papers began to build on the idea and innovate to address issues of low resolution and high computational cost. And, while performance continued to improve with each new method, these papers seemed to focus solely on the design of the transformer encoder and neglected the decoder. Enter SegFormer. SegFormer sets itself apart with:

• a new "positional-encoding-free and hierarchical Transformer encoder"
• "a lightweight All-MLP decoder design"

The novel encoder is able operate at arbitrary resolutions without impacting performance. Additionally, the encoder is able to generate both high resolution and low resolution features in contrast to ViT. The decoder design is able to combine both local and global attention to produce high quality representations at low cost.

With these novel improvements, SegFormer sets a new SOTA on ADE20K, Cityscapes, and COCO-Stuff semantic segmentation datasets.