Research map / Learning theory and algorithms / Optimal transport
Annotation-Efficient Coreset Selection for Context-dependent Segmentation
Jin Zhang, Zhe Cao, Biwen Yang, Ruiheng Zhang
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 |
|---|---|---|---|
| use point-based optimal-transport cost with MDE instead of fully supervised scoring (TFDP) | CostSet · SOD (DUTS), 80% pruning | IoU: 62.1 -> 66.2 | Improved |
| use CostSet at 40% pruning versus full data | UNet ResNet50 · COD (COD10K) | IoU: 55.1 -> 52.0 | Hurt |
| use CostSet on TOD where gains are smaller | CostSet · TOD (GDD), 80% pruning | IoU: 64.8 -> 65.6 (vs CCS) | Improved |
| add MDE selection instead of TopK on ADP cost (ablation) | CostSet · SOD (DUTS), 80% pruning | IoU: 64.4 -> 66.2 | Improved |
| add MDE selection instead of TailK on ADP cost (ablation) | CostSet · SOD (DUTS), 80% pruning | IoU: 63.1 -> 66.2 | Improved |
| add MDE selection instead of TopK at low pruning (ablation) | CostSet · SOD (DUTS), 10% pruning | IoU: 74.9 -> 74.8 | No clear effect |
| use MDE instead of CCS/Entropy/EL2N/Forgetting/TFDP on same transport cost (ablation) | CostSet · SOD (DUTS), 80% pruning | IoU: 64.3 -> 66.2 (vs CCS) | Improved |
| use transport cost C scoring in other selection methods versus their original scores (ablation) | Entropy/Forgetting/EL2N/CCS/TFDP · SOD (DUTS) | IoU: e.g. CCS +6.3 at 80%, TFDP +0.8 | Improved |
| replace point labels by GT pixel-level cost (ablation) | CostSet · SOD (DUTS) | IoU: 66.2 -> 66.9 (80%) | No clear effect |
| use box or scribble labels instead of points in ADP (ablation) | CostSet · SOD (DUTS) | IoU across pruning rates: label time 2s -> 6s (box) | No clear effect |
| use point labels instead of boxes (ablation) | ADP · SOD (DUTS) | background coverage Bc: 15.7% -> 8.3% | Improved |
| tune drop ratio per task (0.15 for SOD/MIS, 0.1 for others) (ablation) | MDE · six CD tasks | IoU | Mixed |
Closest papers by meaning
- ELFS: Label-Free Coreset Selection with Proxy Training Dynamics (ICLR 2025 · similarity 0.85)
- Spider: A Unified Framework for Context-dependent Concept Segmentation (ICML 2024 · similarity 0.85)
- $\mathbb{D}^2$ Pruning: Message Passing for Balancing Diversity & Difficulty in Data Pruning (ICLR 2024 · similarity 0.84)
- ContextShift: A Controlled Benchmark for Context Dependence in Object Detection (NeurIPS 2026 · similarity 0.84)
- Boundary Matters: A Bi-Level Active Finetuning Method (NeurIPS 2024 · similarity 0.84)
- Data Pruning by Information Maximization (ICLR 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.
