Research map / Detection and segmentation / Weakly supervised detection and segmentation
Weakly Supervised Training of Universal Visual Concepts for Multi-domain Semantic Segmentation
Petra Bevandić, Marin Oršić, Josip Šarić, Ivan Grubišić, Siniša Šegvić
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add universal taxonomy with NLL+ instead of naive concatenation (ablation) | SNp-RN18 · ADE20k val (joint Vistas-ADE20k training) | mIoU: 36.8 -> 37.4 | Improved |
| add universal taxonomy with NLL+ instead of naive concatenation (ablation) | SNp-RN18 · WildDash 2 minival (joint Vistas-WD2 training) | mIoU: 55.3 -> 56.2 | Improved |
| add post-inference mapping to naive concatenation (ablation) | SNp-RN18 · ADE20k val (joint Vistas-ADE20k training), Vistas | mIoU: 41.0 -> 42.2 | Improved |
| add post-inference mapping to naive concatenation (ablation) | SNp-RN18 · Cityscapes val (joint Vistas-City training) | mIoU: 76.8 -> 76.8 | No clear effect |
| train jointly on Vistas and ADE20k with NLL+ instead of single dataset (ablation) | SNp-RN18 · Vistas val (joint Vistas-ADE20k training) | mIoU: 46.4 -> 42.8 | Hurt |
| add NLL+ weak supervision instead of two-step pseudo-label training (ablation) | SNp-RN18 · ADE20k val (joint Vistas-ADE20k training) | mIoU: 34.1 -> 37.4 | Improved |
| add NLL+ instead of NLL-max (max over universal classes instead of sum) (ablation) | SNp-RN18 · Cityscapes relabeled split | mIoU: 61.8 -> 74.3 | Improved |
| add NLL+ instead of naive concatenation (ablation) | SNp-RN18 · Cityscapes relabeled split | mIoU: 67.6 -> 74.3 | Improved |
| add NLL+ instead of partial merge (ablation) | SNp-RN18 · Cityscapes relabeled split | mIoU: 71.4 -> 74.3 | Improved |
| replace ResNet-18 pixel-level model by Mask2Former with universal taxonomy (ablation) | NLL+ universal · Cross-dataset WD2 | mIoU: 43.9 -> 49.8 | Improved |
| add universal taxonomy instead of MSeg relabeling (ablation) | SNp-DN161 · Vistas val (MSeg collection, default protocol) | mIoU: 28.7 -> 44.2 | Improved |
| add universal taxonomy instead of MSeg relabeling (ablation) | SNp-DN161 · Cityscapes val (MSeg collection, default protocol) | mIoU: 77.0 -> 76.1 | Hurt |
Closest papers by meaning
- Weakly Supervised Semantic Segmentation for Driving Scenes (AAAI 2024 · similarity 0.89)
- Automated Label Unification for Multi-Dataset Semantic Segmentation with GNNs (NeurIPS 2024 · similarity 0.87)
- UniDSeg: Unified Cross-Domain 3D Semantic Segmentation via Visual Foundation Models Prior (NeurIPS 2024 · similarity 0.86)
- DynAlign: Unsupervised Dynamic Taxonomy Alignment for Cross-Domain Segmentation (ICLR 2025 · similarity 0.86)
- Leveraging Depth and Language for Open-Vocabulary Domain-Generalized Semantic Segmentation (NeurIPS 2025 · similarity 0.86)
- Stronger Fewer & Superior: Harnessing Vision Foundation Models for Domain Generalized 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.
