Reproduce “Unsupervised Semantic Segmentation Facilitates Model Understanding” 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
Xiaoyan Yu, Lisa Mais, Jannik Franzen, Peter Hirsch, Nick Lechtenbörger, Andreas Mardt, Dagmar Kainmüller
The authors compare frozen vision-transformer features by clustering patches across images, revealing whether assignments follow image content or recurring locations. Across eight self-supervised models and two baselines, positional effect is negatively correlated with segmentation mean intersection-over-union (Spearman ρ = −0.59, p < .001). DINOv3-Large shows stronger positional effects and lower mIoU than Base, but this association does not establish causality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.