Reproduce “OPTIMUS: Observing Persistent Transformations in Multi-temporal Unlabeled Satellite-data” 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
Raymond Yu, Paul Han, Josh Myers-Dean, Piper Wolters, Favyen Bastani
Researchers from the University of Washington and the Allen Institute for AI developed OPTIMUS, a self-supervised learning method that identifies persistent, long-lasting changes in satellite imagery by learning relative temporal ordering from unlabeled data. It achieved an 87.6% AUROC for detecting changes in 512x512 image time series and 94.15% AUROC for localizing changes in 128x128 patches, outperforming existing unsupervised baselines.
Key insights
- Learning the long-term temporal ordering of satellite images implicitly enables the model to identify persistent changes while being robust to transient seasonal variations.
- Designing robust anchor sets (multiple consecutive images) and strategic query positioning in self-supervised training is crucial for mitigating image quality issues and conflicting signals.
- Domain-specific pre-training (Satlas weights) and an iterative training strategy for localization significantly enhance both overall change detection and fine-grained spatial accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.