MemOVCD: Training-Free Open-Vocabulary Change Detection Framework
MemOVCD is a training-free open-vocabulary change detection framework for bi-temporal remote sensing images, introduced in arXiv paper 2604.26774. It addresses limitations of existing methods that process each timestamp independently or only interact at the final comparison stage, leading to insufficient temporal coupling and fragmented change regions. The framework reformulates change detection as a two-frame tracking problem, using cross-temporal memory reasoning and global-local adaptive rectification. It leverages foundation models like SAM, DINO, and CLIP without requiring training. The approach aims to distinguish genuine semantic changes from non-semantic appearance discrepancies and maintain global semantic continuity in high-resolution images.
Key facts
- MemOVCD is a training-free framework for open-vocabulary change detection.
- It uses cross-temporal memory reasoning and global-local adaptive rectification.
- The framework reformulates bi-temporal change detection as a two-frame tracking problem.
- It leverages foundation models SAM, DINO, and CLIP.
- Existing methods process timestamps independently or interact only at the final stage.
- The approach addresses insufficient temporal coupling and fragmented change regions.
- It targets bi-temporal remote sensing images.
- The paper is available on arXiv with ID 2604.26774.
Entities
Institutions
- arXiv