Making Vision Transformers Truly Shift-Equivariant

TL;DR

This paper introduces data-adaptive modules for Vision Transformers to achieve true shift-equivariance, ensuring consistent outputs under input shifts while maintaining competitive performance on classification and segmentation tasks.

Contribution

The paper proposes novel, data-adaptive modules for ViTs that enable shift-equivariance, a property lacking in traditional ViTs, across multiple architectures.

Findings

Achieves 100% shift consistency on four ViT models.

Maintains competitive accuracy on image classification tasks.

Performs well on semantic segmentation across datasets.

Abstract

For computer vision, Vision Transformers (ViTs) have become one of the go-to deep net architectures. Despite being inspired by Convolutional Neural Networks (CNNs), ViTs' output remains sensitive to small spatial shifts in the input, i.e., not shift invariant. To address this shortcoming, we introduce novel data-adaptive designs for each of the modules in ViTs, such as tokenization, self-attention, patch merging, and positional encoding. With our proposed modules, we achieve true shift-equivariance on four well-established ViTs, namely, Swin, SwinV2, CvT, and MViTv2. Empirically, we evaluate the proposed adaptive models on image classification and semantic segmentation tasks. These models achieve competitive performance across three different datasets while maintaining 100% shift consistency.