Reproduce “SAM-Aug: Leveraging SAM Priors for Few-Shot Parcel Segmentation in Satellite Time Series” 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
Kai Hu, Yaozu Feng, Vladimir Lysenko, Ya Guo, Huayi Wu
The SAM-Aug framework enhances few-shot parcel segmentation in satellite time series by integrating geometry-aware mask priors from the Segment Anything Model (SAM) with a novel temporal consistency loss. This approach yielded an average test mIoU of 36.21%, an absolute improvement of +2.33% over baseline models, in settings with only 5% labeled data.
Key insights
- Geometry-aware priors from a foundation model like SAM, even if semantic-agnostic, can serve as an effective inductive bias to regularize few-shot learning in a different domain (remote sensing).
- Enforcing temporal consistency of predictions within these pre-defined spatial regions significantly reduces overfitting to sparse labels and improves model generalization in low-data regimes.
- SAM-Aug offers a 'plug-and-play' approach that leverages the powerful segmentation capabilities of foundation models without requiring additional manual annotations or complex model adaptation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.