Reproduce “SeasonStereo: Robust Dense Stereo Matching for Multi-Date Satellite Imagery via Generative AI” 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
Álvaro Díaz-Laureano, Roger Marí, Elías Masquil, Pablo Arias, Gabriele Facciolo
SeasonStereo enables robust dense stereo matching from multi-date satellite imagery by leveraging generative AI for geometry-preserving synthetic data augmentation and fine-tuning a foundation stereo model with a scalable multi-term supervision loss. The framework achieved an average altitude MAE of 1.05 meters, matching LiDAR-supervised methods, and yielded sharper geometric details on building contours under diverse seasonal conditions.
Key insights
- Generative AI can effectively synthesize realistic seasonal variations while preserving geometric fidelity, enabling scalable data augmentation to overcome scarcity of diachronic training data in remote sensing.
- Combining zero-shot geometric priors from powerful foundation models with photometric consistency on original synchronic views (rather than diachronic inputs) significantly improves fine-grained structural detail.
- Curbing loss computation to stable regions like buildings and excluding inconsistent synthetic generations (e.g., water, trees) is crucial for maintaining geometric accuracy when using generative models for data augmentation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.