Making Vision Transformers Truly Shift-Equivariant
This work addresses a key limitation in ViTs for computer vision applications, offering a solution to improve robustness to spatial shifts, though it is incremental as it builds on existing ViT frameworks.
The paper tackled the problem of Vision Transformers (ViTs) lacking shift invariance by introducing novel data-adaptive designs for modules like tokenization and self-attention, achieving 100% shift consistency on four ViT architectures while maintaining competitive performance on image classification and semantic segmentation tasks across three datasets.
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.