Research map / Label-efficient and robust learning / Semi-supervised learning
Jaccard Metric Losses: Optimizing the Jaccard Index with Soft Labels
Zifu Wang, Xuefei Ning, Matthew B. Blaschko
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 |
|---|---|---|---|
| Replace CE-only training with JML plus CE | DL3-R18 segmentation · Cityscapes | mIoU: 72.88±0.44 -> 75.55±0.13 | Improved |
| Replace CE-only training with JML plus CE | DL3-R18 segmentation · PASCAL VOC | mIoU: 72.47±0.33 -> 74.42±0.52 | Improved |
| Add boundary label smoothing to JML | DL3-R18 segmentation · PASCAL VOC | mIoU: 74.42±0.52 -> 75.60±0.24 | Improved |
| Add JML-based teacher distillation | DL3-R18 segmentation · Cityscapes | mIoU: 75.55±0.13 -> 76.68±0.33 | Improved |
| Use JML-based SSL with boundary label smoothing | AugSeg, DL3+-R101 · PASCAL VOC 92-label split | mIoU: 71.09 -> 76.43 | Improved |
| Set CE/JML loss weights to 0.25/0.75 rather than CE-only (ablation) | DL3-R101 JML · PASCAL VOC | mIoU: 78.39±0.09 -> 80.26±0.45 | Improved |
| Set CE/JML loss weights to 0.50/0.50 instead of 0.25/0.75 (ablation) | DL3-R101 JML · PASCAL VOC | mIoU: 80.26±0.45 -> 79.92±0.12 | Hurt |
| Add the teacher's CE and JML distillation loss terms (ablation) | DL3-R18 student with supervised JML · Cityscapes | mIoU: 75.55±0.13 -> 76.16±0.37 | Improved |
| Train the teacher with JML-BLS (ablation) | DL3-R18 JML-KD · PASCAL VOC | mIoU: 75.05±0.31 -> 75.89±0.29 | Improved |
| Filter classes based on teacher confidence instead of including all classes (ablation) | DL3-R18 JML-KD · PASCAL VOC | mIoU: 75.21±0.33 -> 75.89±0.29 | Improved |
| Use only the teacher's maximum-confidence class rather than confidence-filtered classes (ablation) | DL3-R18 JML-KD · PASCAL VOC | mIoU: 75.89±0.29 -> 74.50±0.25 | Hurt |
| Add supervised JML loss to the supervised-plus-unlabeled CE-SSL objective (ablation) | DL3+-R101 JML-SSL · PASCAL VOC 92-label split | mIoU: 71.09 -> 72.89 | Improved |
Closest papers by meaning
- The Lovász-Softmax Loss: A Tractable Surrogate for the Optimization of the Intersection-Over-Union Measure in Neural Networks (CVPR 2018 · similarity 0.92)
- Distribution-Aware Margin Calibration for Semantic Segmentation in Images (IJCV 2021 · similarity 0.89)
- Revisiting Evaluation Metrics for Semantic Segmentation: Optimization and Evaluation of Fine-grained Intersection over Union (NeurIPS 2023 · similarity 0.87)
- Adaptive Spatial-BCE Loss for Weakly Supervised Semantic Segmentation (ECCV 2022 · similarity 0.86)
- Complex Wavelet Mutual Information Loss: A Multi-Scale Loss Function for Semantic Segmentation (ICML 2025 · similarity 0.84)
- Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity (CVPR 2025 · similarity 0.84)
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.
