Multiscale Vision Transformers
This work addresses efficient and effective visual recognition for computer vision applications, representing an incremental improvement by integrating multiscale hierarchies into transformers.
The authors tackled the problem of video and image recognition by introducing Multiscale Vision Transformers (MViT), which connect multiscale feature hierarchies with transformer models, resulting in outperforming concurrent vision transformers that are 5-10x more costly in computation and parameters.
We present Multiscale Vision Transformers (MViT) for video and image recognition, by connecting the seminal idea of multiscale feature hierarchies with transformer models. Multiscale Transformers have several channel-resolution scale stages. Starting from the input resolution and a small channel dimension, the stages hierarchically expand the channel capacity while reducing the spatial resolution. This creates a multiscale pyramid of features with early layers operating at high spatial resolution to model simple low-level visual information, and deeper layers at spatially coarse, but complex, high-dimensional features. We evaluate this fundamental architectural prior for modeling the dense nature of visual signals for a variety of video recognition tasks where it outperforms concurrent vision transformers that rely on large scale external pre-training and are 5-10x more costly in computation and parameters. We further remove the temporal dimension and apply our model for image classification where it outperforms prior work on vision transformers. Code is available at: https://github.com/facebookresearch/SlowFast