Research map / Detection and segmentation / Open-set and few-shot detection
Flattening the Parent Bias: Hierarchical Semantic Segmentation in the Poincare Ball
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that flat segmentation networks with bottom-up parent inference generalize better than hierarchical HSSN across domains, and that flat classifiers embedded in the Poincaré ball reduce parent bias and improve parent-level accuracy and calibration.
Problem
Hierarchical semantic segmentation methods report gains on in-domain data, but it is unclear whether the gains come from semantic hierarchy and how well they generalize to unseen domains; flat Euclidean classifiers may show a parent bias in parent-level predictions.
Approach
Flat classifiers infer parent posteriors from child posteriors (Proposition 1), and embedding pixel features in the Poincaré ball makes inter-class hyperbolic distances more uniform (concavity of hyperbolic distance), removing the Euclidean parent bias.
Key results
- On the hardest domains (BDD/Wilddash) Flat-Hyp improves parent-level mIoU over Flat-Euc by +0.22/+3.55 with DeepLabV3+.
- Flat-Hyp lowers parent-level cwECE versus HSSN on BDD and Wilddash by -2.27 and -5.43 with DeepLabV3+.
- Flat classifiers consistently outperform the hierarchical HSSN model on the unseen test domains, and the non-semantic hierarchies yield nearly identical accuracy to semantic ones.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace hierarchical HSSN training with flat Euclidean classification | DeepLabV3+/ResNet-101 · Mapillary, Level 0 | mIoU: 59.34 -> 63.94 | Helped |
| Replace hierarchical HSSN training with flat Euclidean classification | DeepLabV3+/ResNet-101 · BDD, Level 1 | mIoU: 74.28 -> 76.27 | Helped |
| Use flat hyperbolic rather than flat Euclidean classification | DeepLabV3+/ResNet-101 · Wilddash, Level 1 | mIoU: 58.98 -> 62.53 | Helped |
| Use flat hyperbolic rather than flat Euclidean classification | DeepLabV3+/ResNet-101 · ACDC, Level 0 | mIoU: 54.34 -> 47.72 | Hurt |
| Use flat hyperbolic rather than flat Euclidean classification | DeepLabV3+/ResNet-101 · BDD, Level 1 | mIoU: 76.27 -> 76.49 | Helped |
| Use flat hyperbolic rather than flat Euclidean classification | DeepLabV3+/ResNet-101 · BDD, Level 0 | mIoU: 51.54 -> 49.64 | Hurt |
| Use flat hyperbolic rather than flat Euclidean classification | DeepLabV3+/ResNet-101 · Wilddash, Level 0 | mIoU: 39.38 -> 40.57 | Helped |
| Train with semantically meaningless rather than semantic parent groupings (ablation) | HSSN-style DeepLabV3+ · Cityscapes val | mIoU: 80.28 -> 80.2 -> 79.86 | No clear effect |
| Train with semantically meaningless rather than semantic parent groupings (ablation) | HSSN-style DeepLabV3+ · Cityscapes val | mAcc: 86.27 -> 86.18 -> 85.69 | No clear effect |
| Train with semantically meaningless rather than semantic parent groupings (ablation) | HSSN-style DeepLabV3+ · Cityscapes val | aAcc: 96.01 -> 95.92 -> 95.9 | No clear effect |
Surprising findings
- Semantically meaningless meta-class hierarchies yield the same leaf-level segmentation accuracy as the semantic hierarchy.
- A flat classifier outperforms the hierarchical HSSN approach even on parent-level predictions.
Limitations
- Evaluation limited to traffic-scene datasets due to lack of datasets with other taxonomies.
- Analysis restricted to the Poincaré ball; other hyperbolic models such as the Lorentz model are not studied.
- Flat-Hyp does not consistently beat Flat-Euc on parent mIoU in less challenging domains.
Experimental setup
Datasets: Cityscapes
Benchmarks: Cityscapes Mapillary IDD ACDC BDD Wilddash
Metrics: mIoU mAcc aAcc cwECE
Baselines: HSSN LogicSeg Flat-Euc
Builds on: HSSN Atigh et al. hyperbolic image segmentation Ganea et al. hyperbolic neural networks DeepLabV3+ OCRNet
Closest papers by meaning
- Hierarchy-Agnostic Unsupervised Segmentation: Parsing Semantic Image Structure (NeurIPS 2024 · similarity 0.90)
- Learning Hierarchical Image Segmentation For Recognition and By Recognition (ICLR 2024 · similarity 0.89)
- Understanding Fine-tuning CLIP for Open-vocabulary Semantic Segmentation in Hyperbolic Space (CVPR 2025 · similarity 0.87)
- Label-Efficient Few-Shot Semantic Segmentation with Unsupervised Meta-Training (AAAI 2024 · similarity 0.86)
- Diffusion-Driven Two-Stage Active Learning for Low-Budget Semantic Segmentation (NeurIPS 2025 · similarity 0.86)
- SED: A Simple Encoder-Decoder for Open-Vocabulary Semantic Segmentation (CVPR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.