Vision Transformer Adapter for Dense Predictions

This work investigates a simple yet powerful adapter for Vision Transformer (ViT). Unlike recent visual transformers that introduce vision-specific inductive biases into their architectures, ViT achieves inferior performance on dense prediction tasks due to lacking prior information of images. To solve this issue, we propose a Vision Transformer Adapter (ViT-Adapter), which can remedy the defects of ViT and achieve comparable performance to vision-specific models by introducing inductive biases via an additional architecture. Specifically, the backbone in our framework is a vanilla transformer that can be pre-trained with multi-modal data. When fine-tuning on downstream tasks, a modality-specific adapter is used to introduce the data and tasks' prior information into the model, making it suitable for these tasks. We verify the effectiveness of our ViT-Adapter on multiple downstream tasks, including object detection, instance segmentation, and semantic segmentation. Notably, when using HTC++, our ViT-Adapter-L yields 60.1 box AP and 52.1 mask AP on COCO test-dev, surpassing Swin-L by 1.4 box AP and 1.0 mask AP. For semantic segmentation, our ViT-Adapter-L establishes a new state-of-the-art of 60.5 mIoU on ADE20K val, 0.6 points higher than SwinV2-G. We hope that the proposed ViT-Adapter could serve as an alternative for vision-specific transformers and facilitate future research.