Vision Transformers Need Registers
This addresses a problem for researchers and practitioners using Vision Transformers in computer vision, improving feature map quality and performance, though it is incremental as it builds on existing ViT architectures.
The paper identifies artifacts in Vision Transformer feature maps during inference, where high-norm tokens appear in low-informative background areas, and proposes adding extra input tokens to fix this issue, leading to smoother feature maps and setting a new state of the art for self-supervised models on dense visual prediction tasks.
Transformers have recently emerged as a powerful tool for learning visual representations. In this paper, we identify and characterize artifacts in feature maps of both supervised and self-supervised ViT networks. The artifacts correspond to high-norm tokens appearing during inference primarily in low-informative background areas of images, that are repurposed for internal computations. We propose a simple yet effective solution based on providing additional tokens to the input sequence of the Vision Transformer to fill that role. We show that this solution fixes that problem entirely for both supervised and self-supervised models, sets a new state of the art for self-supervised visual models on dense visual prediction tasks, enables object discovery methods with larger models, and most importantly leads to smoother feature maps and attention maps for downstream visual processing.