Co-Scale Conv-Attentional Image Transformers

TL;DR

Co-Scale Conv-Attentional Image Transformers (CoaT) introduce co-scale and conv-attentional mechanisms to enhance multi-scale and contextual modeling in image classification, achieving superior results on ImageNet and downstream tasks.

Contribution

The paper proposes novel co-scale and conv-attentional mechanisms that improve multi-scale and contextual understanding in Transformer-based image classifiers.

Findings

CoaT models outperform similar-sized CNNs and Transformers on ImageNet.

CoaT's backbone improves object detection and segmentation performance.

Efficient multi-scale and contextual modeling enhances classification accuracy.

Abstract

In this paper, we present Co-scale conv-attentional image Transformers (CoaT), a Transformer-based image classifier equipped with co-scale and conv-attentional mechanisms. First, the co-scale mechanism maintains the integrity of Transformers' encoder branches at individual scales, while allowing representations learned at different scales to effectively communicate with each other; we design a series of serial and parallel blocks to realize the co-scale mechanism. Second, we devise a conv-attentional mechanism by realizing a relative position embedding formulation in the factorized attention module with an efficient convolution-like implementation. CoaT empowers image Transformers with enriched multi-scale and contextual modeling capabilities. On ImageNet, relatively small CoaT models attain superior classification results compared with similar-sized convolutional neural networks and image/vision Transformers. The effectiveness of CoaT's backbone is also illustrated on object detection and instance segmentation, demonstrating its applicability to downstream computer vision tasks.