Reproduce “When GeoFMs Meet VHR: Benchmarking Foundation Models on Very High Resolution Earth Observation 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
Kshitij Raj Sharma, Pierre Adorni, Minh-Tan Pham, Dirk Tiede, Sébastien Lefèvre
The authors introduce VHR-Bench to evaluate pretrained visual encoders on very high-resolution RGB satellite and aerial tasks. The benchmark tests twelve encoders with frozen backbones and limited or full labels. Continuing DINOv3 ViT-L on unlabeled VHR imagery raises its aggregate score from 64.0 to 65.2 at 10% labels and from 69.5 to 70.0 at full labels, without changing inference parameters or computation. These results apply to the benchmark’s frozen-backbone protocol.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.