MSViT: Dynamic Mixed-scale Tokenization for Vision Transformers | IEEE Conference Publication | IEEE Xplore