AutoFocusFormer Provides Differentiable Token Downsampling

Kosta Derpanis, Associate Professor at York University specializing in computer vision and deep learning with affiliations at Samsung AI and Vector Institute, retweeted @fuxinli2. The post states that the authors' AutoFocusFormer from CVPR 2023 supplies a differentiable manner to learn to downsample to the most important tokens in an image. The statement is presented as first-party description of the paper's contribution.