Research map / Detection and segmentation / Weakly supervised detection and segmentation
Novel Class Discovery in Semantic Segmentation
Yuyang Zhao, Zhun Zhong, Nicu Sebe, Gim Hee Lee
Check 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 |
|---|---|---|---|
| use over-clustering instead of exact-clustering (ablation) | Basic framework · PASCAL-5^i | novel-class mIoU: 4.7% improvement in average mIoU; Fold0 1.6%; Fold1 -1.88%; Fold2 9.00%; Fold3 10.09% | Mixed |
| add entropy ranking and train with the clean split (ablation) | Basic with over-clustering · PASCAL-5^i | novel-class mIoU (AVG): 54.51 vs 54.36 | No clear effect |
| add dynamic reassignment of clean data (ablation) | Over-clustering + entropy ranking · PASCAL-5^i | novel-class mIoU (AVG): 0.78% improvement in average mIoU | Improved |
| add self-training on unclean data (ablation) | Over-clustering + entropy ranking · PASCAL-5^i | novel-class mIoU (AVG): 2.3% improvement in average mIoU | Improved |
| combine over-clustering, entropy ranking, dynamic reassignment, and self-training (ablation) | Basic framework · PASCAL-5^i | novel-class mIoU (AVG): outperforming the basic framework by 9.28% | Improved |
| use over-clustering (ablation) | Basic framework · COCO-20^i | novel-class mIoU: Fold0 14.81%; Fold2 2.59% | Mixed |
| add the EUMS framework (ablation) | Basic framework · COCO-20^i | novel-class mIoU (AVG): outperforms the baseline by 6.9% | Improved |
| set the clean split ratio too low (0.33 or 0.50) (ablation) | EUMS · PASCAL-5^i | novel-class mIoU | Hurt |
| increase the clean split ratio beyond 0.67 (ablation) | EUMS · PASCAL-5^i | novel-class mIoU | Hurt |
| raise the online pseudo-label threshold to 0.9 (ablation) | EUMS self-training · PASCAL-5^i | novel-class mIoU: 3% | Hurt |
| vary the online pseudo-label threshold within η < 0.5 (ablation) | EUMS self-training · PASCAL-5^i | novel-class mIoU | No clear effect |
| remove self-training by setting η = 1.0 (ablation) | EUMS · PASCAL-5^i | novel-class mIoU | Hurt |
Closest papers by meaning
- Novel Class Discovery for 3D Point Cloud Semantic Segmentation (CVPR 2023 · similarity 0.90)
- A Unified Objective for Novel Class Discovery (ICCV 2021 · similarity 0.89)
- Novel Class Discovery Meets Foundation Models for 3D Semantic Segmentation (IJCV 2024 · similarity 0.89)
- Class Distribution-induced Attention Map for Open-vocabulary Semantic Segmentations (ICLR 2025 · similarity 0.89)
- Cs2K: Class-specific and Class-shared Knowledge Guidance for Incremental Semantic Segmentation (ECCV 2024 · similarity 0.89)
- Semantic-Guided Novel Category Discovery (AAAI 2024 · similarity 0.88)
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.
