Computer Vision News - November 2021

3 Summary 1 Han Hu Best of ICCV 2021 “ We are at the start of an exciting new era for computer vision, ” he tells us. “ Before this paper, I heard different voices about the future of Transformers in computer vision. Some said that Transformer performs well for some vision tasks but has a long way to go to be really popular in computer vision. Others said Transformer has a lot of potential and it should get more attention from the community. We are definitely in the second camp! ” To prove that point, Han and his team showed Swin Transformer achieved strong performance on two important computer vision benchmarks: COCO object detection and ADE20K semantic segmentation , setting a new record on COCO and ADE20K. They demonstrated that the Transformer backbone can be around three points better than previous CNN backbones. Top: from NLP Transformers to Swin Transformers (vision) Bottom: a key design towards fast local Transformer computation by non-overlapping shifted window