Reproduce “ChangeFlow -- Latent Rectified Flow for Change Detection in Remote Sensing” on OpenResearch

About this paper

Blaž Rolih, Matic Fučka, Filip Wolf, Luka Čehovin Zajc

ChangeFlow introduces a generative framework for remote sensing change detection that synthesizes change masks in a compact latent space using rectified flow. The method achieves state-of-the-art performance on multiple binary and semantic change detection benchmarks, producing more spatially coherent and consistent masks with improved computational efficiency.

Key insights

• Generative modeling in a compact latent space, specifically utilizing rectified flow, can overcome the efficiency and performance limitations of prior pixel-space generative methods for RSCD.
• A simple, parameter-free conditioning signal based on bi-temporal feature differences is effective and robust for guiding latent generative models, outperforming more complex conditioning architectures.
• Directly modeling change masks as coherent objects through a generative process inherently yields spatially consistent predictions with fewer spurious holes and better-defined boundaries than pixel-wise discriminative approaches.

Summary by alphaXiv.