Reproduce “Lens: Training-Free Semantic Search for Cross-Tile Object and Scene Retrieval in Satellite Imagery” 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

Ata Akbari Asanjan, Anit Upadhyaya, Jeremy Kravitz, Subash Yeggina

Lens lets analysts retrieve similar patches across an indexed satellite area by clicking an example, without training a class-specific detector. It uses frozen DINOv3 features, removing positional bias and centering them around the area’s mean. On DFC2020, Lens achieves mean pixel F1 of 0.385 across eight land-cover classes, beating pairwise INSID3 on six. This measures agreement with land-cover labels, not object-detection accuracy.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.