Reproduce “LG-CD: Enhancing Language-Guided Change Detection through SAM2 Adaptation” 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
Yixiao Liu, Yizhou Yang, Jinwen Li, Jun Tao, Ruoyu Li, Xiangkun Wang, Min Zhu, Junlong Cheng
A language-guided change detection (LG-CD) model, developed at Sichuan University, integrates a SAM2-based encoder with explicit text fusion and visual-semantic decoding to enhance remote sensing change detection. The model achieved F1-scores of 90.35% on LEVIR-CD and 91.83% on WHU-CD, allowing for targeted detection guided by natural language prompts.
Key insights
- Integrating natural language descriptions into change detection tasks provides valuable semantic context, leading to a more comprehensive understanding of changes beyond unimodal visual input.
- Adapting a powerful pre-trained vision foundation model (SAM2) with lightweight layers is an effective strategy for feature extraction and generalization in remote sensing change detection.
- Explicit attention mechanisms for visual-textual fusion enable the model to interpret language cues, allowing for a semantically-guided focus on specific types of changes, improving detection accuracy and reducing false positives/negatives.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.