Reproduce “Flattening the Parent Bias: Hierarchical Semantic Segmentation in the Poincaré Ball” 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
Simon Weber, Barış Zöngür, Nikita Araslanov, Daniel Cremers
This research re-examines the role of semantic hierarchies in segmentation, finding that reported performance gains may not be tied to semantic meaning. It proposes a hyperbolic embedding approach in the Poincar´e ball model for hierarchical semantic segmentation, which improves generalization and calibration for parent-level predictions. The hyperbolic model (Flat-Hyp) achieved a +3.55 mIoU increase for Level 1 over Euclidean flat classifiers on challenging out-of-domain datasets like Wilddash.
Key insights
- The benefits of hierarchical semantic segmentation, as reported in prior work, may not be inherently tied to the semantic meaning of the label hierarchy itself.
- Simple flat classifiers, which infer parent categories from child predictions, serve as robust and often superior baselines, especially in cross-domain generalization.
- Hyperbolic geometry, specifically the Poincar´e ball model, can effectively "flatten" the parent bias observed in Euclidean spaces, leading to more uniform inter-class relationships and improved model calibration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.