Multiscale Vision Transformers

TL;DR

Multiscale Vision Transformers (MViT) introduce a hierarchical multiscale architecture for video and image recognition, outperforming existing transformers in accuracy and efficiency by leveraging multiscale feature hierarchies.

Contribution

The paper proposes a novel multiscale transformer architecture that hierarchically expands channel capacity while reducing spatial resolution, improving recognition performance without large external pre-training.

Findings

Outperforms concurrent vision transformers in video recognition tasks.

Achieves higher accuracy on image classification benchmarks.

Requires significantly less computation and fewer parameters.

Abstract

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