Reproduce “Show Me What and Where has Changed? Question Answering and Grounding for Remote Sensing 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
Ke Li, Fuyu Dong, Di Wang, Shaofeng Li, Quan Wang, Xinbo Gao, Tat-Seng Chua
Xidian University researchers introduce Change Detection Question Answering and Grounding (CDQAG), a new task for remote sensing that concurrently provides textual answers and pixel-level visual evidence to user questions about changes in multi-temporal images. They also present QAG-360K, the first large-scale benchmark dataset for this task, along with VisTA, a baseline model achieving 73.03% Average Accuracy and 44.94% Mean IoU on the new dataset.
Key insights
- Simultaneously providing textual answers and pixel-level visual evidence for change detection significantly enhances interpretability and user confidence in automated systems.
- Designing a benchmark with diverse question types and challenging attributes like temporal semantic correlation and fragmented masks is essential for developing robust and precise CDQAG models.
- A multi-stage architecture leveraging language-guided feature aggregation and cross-modal reasoning, along with contrastive learning, can effectively unify question answering and visual grounding for complex remote sensing change detection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.