Reproduce “ViLaCD-R1: A Vision-Language Framework for Semantic Change Detection in Remote Sensing” 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
Xingwei Ma, Shiyang Feng, Bo Zhang, Bin Wang
ViLaCD-R1 introduces a two-stage vision-language framework for remote sensing change detection, combining a fine-tuned Vision-Language Model for semantic reasoning with a Mask-Guided Decoder for precise pixel-level localization. This framework achieved state-of-the-art accuracy on LEVIR-CD, LEVIR-CD+, and SYSU-CD datasets, including an IoU of 84.33% on LEVIR-CD.
Key insights
- Fusing VLM-based semantic reasoning with a dedicated pixel-level decoder effectively overcomes the limitations of each approach, leading to both high semantic understanding and precise localization.
- GRPO-based reinforcement learning significantly enhances the VLM's ability to consistently identify genuine semantic changes and suppress non-semantic noise, as demonstrated by noticeable performance drops when omitted.
- The coarse semantic mask generated by the VLM provides a crucial prior for the decoder, dramatically reducing false alarms and guiding fine-grained predictions, proving indispensable for accurate boundary reconstruction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.