Reproduce “ChangeViT: Unleashing Plain Vision Transformers for Change Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Duowang Zhu, Xiaohu Huang, Haiyan Huang, Zhenfeng Shao, Qimin Cheng
ChangeViT introduces a dual-module framework that effectively leverages plain Vision Transformers for change detection by augmenting their global reasoning with a dedicated mechanism for capturing fine-grained local details. The approach achieves state-of-the-art performance on multiple remote sensing datasets, outperforming both CNN-based and complex hierarchical transformer models in detecting multi-scale changes.
Key insights
- Plain Vision Transformers possess an inherent strength in detecting large-scale changes due to their global receptive fields.
- The effectiveness of ViTs in change detection is significantly enhanced by strategically combining their global understanding with auxiliary modules designed for fine-grained local detail capture.
- Utilizing large-scale self-supervised pre-trained weights, such as those from DINOv2, provides substantial performance improvements for ViT-based change detection models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.