Generative Change Detection Model Outperforms Discriminative Methods
A new generative approach for remote sensing change detection (RSCD) called ChangeFlow uses latent rectified flow to model plausible change masks. Unlike per-pixel discriminative classifiers that produce single predictions, ChangeFlow samples from a distribution to capture ambiguity and enforce global consistency. The method addresses high computational costs of pixel-space generation and complex conditioning mechanisms that hindered prior generative RSCD models. ChangeFlow achieves state-of-the-art results on benchmark datasets by operating in a latent space. The paper is available on arXiv under ID 2605.15375.
Key facts
- ChangeFlow is a generative model for remote sensing change detection.
- It uses latent rectified flow to model distributions of change masks.
- Generative approach captures ambiguity and encourages global consistency.
- Prior generative RSCD methods lagged behind discriminative baselines.
- ChangeFlow addresses computational cost and conditioning complexity.
- The method achieves state-of-the-art results on benchmarks.
- Paper available at arXiv:2605.15375.
- Announce type: cross.
Entities
Institutions
- arXiv