In this paper, we study Multiscale Vision Transformers (MViTv2) as a unified architecture for image and video classification, as well as object detection. We present an improved version of MViT that incorporates decomposed relative positional embeddings and residual pooling connections. We instantiate this architecture in five sizes and evaluate it for ImageNet classification, COCO detection and Kinetics video recognition where it outperforms prior work. We further compare MViTv2s' pooling attention to window attention mechanisms where it outperforms the latter in accuracy/compute. Without bells-and-whistles, MViTv2 has state-of-the-art performance in 3 domains: 88.8% accuracy on ImageNet classification, 58.7 boxAP on COCO object detection as well as 86.1% on Kinetics-400 video classification. Code and models are available at https://github.com/facebookresearch/mvit.
@article{arxiv.2112.01526,
title = {MViTv2: Improved Multiscale Vision Transformers for Classification and Detection},
author = {Yanghao Li and Chao-Yuan Wu and Haoqi Fan and Karttikeya Mangalam and Bo Xiong and Jitendra Malik and Christoph Feichtenhofer},
journal= {arXiv preprint arXiv:2112.01526},
year = {2022}
}