Research map / Learning theory and algorithms / Combinatorial and search optimization
Discriminative Training of Conditional Random Fields with Probably Submodular Constraints
Maxim Berman, 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 |
|---|---|---|---|
| relax constraint set from C1 (definitely submodular) to probably submodular C4 (ablation) | probably submodular SSVM · TU Darmstadt cows | global and average accuracy | Improved |
| relax constraints from definitely submodular C2 to probably submodular C4 (ablation) | SSVM · UCSB retinal | IoU (%): 85.0 -> 87.2 | Improved |
| relax constraints from C2 to probably submodular C4 (ablation) | SSVM · UCSB retinal | 1-Delta_avg (%): 95.7 -> 96.6 | Improved |
| use intermediate probably submodular set C3 instead of C2 (ablation) | SSVM · UCSB retinal | IoU (%): 85.0 -> 85.2 | No clear effect |
| add transductive test-set submodularity constraints (C4bar instead of C4) (ablation) | probably submodular SSVM · UCSB retinal | IoU (%): 87.2 -> 87.2 | No clear effect |
| allow pairwise potentials (C1 instead of unary-only C0) (ablation) | SSVM · Mitochondria 3D EM | test IoU (%): 43.9 -> 60.4 | Improved |
| use C3 instead of larger set C4 (ablation) | probably submodular SSVM · Mitochondria 3D EM | test IoU (%): 56.9 -> 64.5 | Mixed |
| train with submodular constraints C4 instead of QPBO approximate inference | SSVM · UCSB retinal | 1-Delta_avg (%): 92.4 -> 96.6 | Improved |
| use constrained C4 with exact graph cuts instead of truncated graph cuts (TGC) | SSVM · scene multi-label | test accuracy (%): 83.3 -> 90.4 | Improved |
| use constrained C4 instead of TGC | SSVM · yeast multi-label | test accuracy (%): 71.8 -> 80.0 | Improved |
| add 2-stage pretraining and delayed constraint generation (ablation) | constraint generation · TU Darmstadt cows | margins computed: 102.5 -> 6.5 (x10^6) | Improved |
| add delayed constraint generation (1-pass) (ablation) | constraint generation · TU Darmstadt cows | total constraint generation time: 400 s -> 329 s | Improved |
Closest papers by meaning
- Blending Learning and Inference in Conditional Random Fields (JMLR 2016 · similarity 0.87)
- Truncated Max-Of-Convex Models (CVPR 2017 · similarity 0.86)
- Joint M-Best-Diverse Labelings as a Parametric Submodular Minimization (NeurIPS 2016 · similarity 0.86)
- Parameter Learning for Log-supermodular Distributions (NeurIPS 2016 · similarity 0.86)
- Approximate Decomposable Submodular Function Minimization for Cardinality-Based Components (NeurIPS 2021 · similarity 0.86)
- Efficient Continuous Relaxations for Dense CRF (ECCV 2016 · 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.
