Research map / Detection and segmentation / CNN features
The Lovász-Softmax Loss: A Tractable Surrogate for the Optimization of the Intersection-Over-Union Measure in Neural Networks
Maxim Berman, Amal Rannen Triki, 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 cross-entropy with Lovász hinge | DeeplabV2 ResNet-101 fine-tuning · Pascal VOC val | image-mIoU: 77.14 -> 80.5 | Improved |
| Replace cross-entropy with Lovász-Softmax | Deeplab-v2 · Pascal VOC test | mIoU (%) multi-scale + CRF: 76.44 -> 79.00 | Improved |
| Replace cross-entropy with Lovász-Softmax fine-tuning | ENet · Cityscapes test | Class IoU: 58.29 -> 63.06 | Improved |
| Fine-tune with Lovász-Softmax instead of cross-entropy | Deeplab-based model · IBSR test, 3-fold mean | Mean Jaccard: 60.16 -> 61.55 | Improved |
| Train Lovász-Softmax for 30K rather than 20K iterations (ablation) | Deeplab-v2 with equibatch · Pascal VOC val | mIoU (%) single-scale: 76.53 -> 77.41 | Improved |
| Add equibatch to Lovász-Softmax training (ablation) | Deeplab-v2 · Pascal VOC val | mIoU (%) multi-scale + CRF: 77.99 -> 78.49 | Mixed |
| Add equibatch to cross-entropy training (ablation) | Deeplab-v2 · Pascal VOC val | mIoU (%) multi-scale + CRF: 76.53 -> 77.31 | Improved |
| Optimize per-image IoU rather than batch/dataset IoU (ablation) | Lovász-Softmax · Pascal VOC | dataset-mIoU | Hurt |
| Average the Lovász-Softmax only over classes present in the batch (ablation) | Lovász-Softmax · Pascal VOC dataset-mIoU | dataset-mIoU | Improved |
| Add Gaussian CRF post-processing (ablation) | Lovász-Softmax Deeplab-v2 · Pascal VOC val | mIoU (%): 78.22 -> 79.12 | Improved |
| Fine-tune ENet with Lovász-Softmax | ENet · Cityscapes test | Class iIoU: 34.36 -> 34.06 | No clear effect |
| Fine-tune ENet with Lovász-Softmax | ENet · Cityscapes test | Cat. iIoU: 63.99 -> 61.05 | Hurt |
Closest papers by meaning
- Jaccard Metric Losses: Optimizing the Jaccard Index with Soft Labels (NeurIPS 2023 · similarity 0.92)
- Distribution-Aware Margin Calibration for Semantic Segmentation in Images (IJCV 2021 · similarity 0.90)
- Revisiting Evaluation Metrics for Semantic Segmentation: Optimization and Evaluation of Fine-grained Intersection over Union (NeurIPS 2023 · similarity 0.88)
- Complex Wavelet Mutual Information Loss: A Multi-Scale Loss Function for Semantic Segmentation (ICML 2025 · similarity 0.86)
- Boundary IoU: Improving Object-Centric Image Segmentation Evaluation (CVPR 2021 · similarity 0.85)
- RankSEG-RMA: An Efficient Segmentation Algorithm via Reciprocal Moment Approximation (NeurIPS 2025 · similarity 0.85)
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.
