Research map / Learning theory and algorithms / Kernels and regression theory
Regularized Frank-Wolfe for Dense CRFs: Generalizing Mean Field and Beyond
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 Euclidean Frank-Wolfe dense CRF | DeepLabv3+ · PASCAL VOC test | mIoU: 87.6 -> 88.0 | Helped |
| Add Euclidean Frank-Wolfe dense CRF | DeepLabv3+ · Cityscapes test | mIoU: 83.5 -> 83.6 | Helped |
| Use Euclidean Frank-Wolfe instead of the strongest reported solver baseline | DeepLabv3+ joint CNN-CRF training · PASCAL VOC val | mIoU: 84.87 -> 85.14 | Helped |
| Use Euclidean Frank-Wolfe instead of the strongest reported solver baseline | DeepLabv3+ joint CNN-CRF training · Cityscapes val | mIoU: 79.80 -> 79.95 | Helped |
| Replace the CRF output with the average of the CRF output and CNN initialization (ablation) | Entropic Frank-Wolfe CRF (λ=0.7) · Joint CNN-CRF training | training stability | Helped |
| Replace the CRF output with the average of the CRF output and CNN initialization (ablation) | Mean field CRF · Joint CNN-CRF training | training performance | Hurt |
| Train the Frank-Wolfe inference updates end-to-end without regularization (ablation) | Vanilla Frank-Wolfe CRF · PASCAL VOC val | validation accuracy | Hurt |
| Make the step size and regularization weight trainable (ablation) | Euclidean and Entropic Frank-Wolfe · Semantic segmentation | mIoU | No clear effect |
| Use entropic Frank-Wolfe with λ=0.3 rather than λ=0.7 (ablation) | Entropic Frank-Wolfe CRF · PASCAL VOC val, DeepLabv3+ joint training | mIoU: 84.64 -> 84.50 | Hurt |
Closest papers by meaning
- Efficient Linear Programming for Dense CRFs (CVPR 2017 · similarity 0.91)
- Gaussian Conditional Random Field Network for Semantic Segmentation (CVPR 2016 · similarity 0.89)
- Boosting the Performance of Generic Deep Neural Network Frameworks with Log-supermodular CRFs (NeurIPS 2022 · similarity 0.88)
- Constraining Linear-chain CRFs to Regular Languages (ICLR 2022 · similarity 0.86)
- Sparse Non-Local CRF (CVPR 2022 · similarity 0.86)
- Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation (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.