ViTAR: Vision Transformer with Any Resolution

TL;DR

ViTAR introduces dynamic resolution adjustment and fuzzy positional encoding to enable Vision Transformers to effectively process images at any resolution, improving scalability and accuracy across diverse image sizes.

Contribution

The paper presents a novel module for dynamic resolution adjustment and fuzzy positional encoding, enhancing ViT's scalability and performance across multiple resolutions.

Findings

Achieves 83.3% top-1 accuracy at 1120x1120 resolution

Maintains 80.4% accuracy at 4032x4032 resolution

Reduces computational costs while improving resolution adaptability

Abstract

This paper tackles a significant challenge faced by Vision Transformers (ViTs): their constrained scalability across different image resolutions. Typically, ViTs experience a performance decline when processing resolutions different from those seen during training. Our work introduces two key innovations to address this issue. Firstly, we propose a novel module for dynamic resolution adjustment, designed with a single Transformer block, specifically to achieve highly efficient incremental token integration. Secondly, we introduce fuzzy positional encoding in the Vision Transformer to provide consistent positional awareness across multiple resolutions, thereby preventing overfitting to any single training resolution. Our resulting model, ViTAR (Vision Transformer with Any Resolution), demonstrates impressive adaptability, achieving 83.3\% top-1 accuracy at a 1120x1120 resolution and 80.4\% accuracy at a 4032x4032 resolution, all while reducing computational costs. ViTAR also shows strong performance in downstream tasks such as instance and semantic segmentation and can easily combined with self-supervised learning techniques like Masked AutoEncoder. Our work provides a cost-effective solution for enhancing the resolution scalability of ViTs, paving the way for more versatile and efficient high-resolution image processing.