Research map / Detection and segmentation / CNN features
Revisiting Evaluation Metrics for Semantic Segmentation: Optimization and Evaluation of Fine-grained Intersection over Union
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 the DeepLabV3+ feature-aggregation branch to DeepLabV3 | DeepLabV3-ResNet101 · Cityscapes | mIoU^C: 69.09±0.06 -> 70.17±0.16 | Helped |
| Replace DeepLabV3-ResNet101 with UNet-ResNet101, which aggregates multi-scale features without an extra receptive-field module | DeepLabV3-ResNet101 · Cityscapes | mIoU^C: 69.09±0.06 -> 68.78±0.09 | No clear effect |
| Use cross-entropy only instead of JML (ablation) | DeepLabV3+-ResNet101 · Cityscapes | mIoU^C: 70.18±0.16 -> 67.61±0.12 | Hurt |
| Use cross-entropy only instead of JML (ablation) | DeepLabV3+-ResNet101 · ADE20K | mIoU^C: 46.52±0.30 -> 39.66±0.13 | Hurt |
| Optimize mIoU^D only rather than the equal mIoU^I/mIoU^C objective (ablation) | DeepLabV3+-ResNet101 · Cityscapes | mIoU^D: 80.91±0.17 -> 81.22±0.27 | Helped |
| Optimize mIoU^D only rather than the equal mIoU^I/mIoU^C objective (ablation) | DeepLabV3+-ResNet101 · Cityscapes | mIoU^C: 70.18±0.16 -> 68.97±0.19 | Hurt |
| Optimize mIoU^C only rather than the equal mIoU^I/mIoU^C objective (ablation) | DeepLabV3+-ResNet101 · Cityscapes | mIoU^C: 70.18±0.16 -> 70.49±0.19 | Helped |
| Optimize mIoU^I only rather than the equal mIoU^I/mIoU^C objective (ablation) | DeepLabV3+-ResNet101 · Cityscapes | mIoU^I: 76.21±0.05 -> 76.28±0.02 | Helped |
| Use a uniform mixture of mIoU^D, mIoU^I, and mIoU^C in JML (ablation) | DeepLabV3+-ResNet101 · Cityscapes | mIoU^C: 70.18±0.16 vs 70.08±0.15 | No clear effect |
| Use a uniform mixture of mIoU^D, mIoU^I, and mIoU^C in JML (ablation) | DeepLabV3+-ResNet101 · ADE20K | mIoU^C: 46.52±0.30 vs 46.77±0.07 | No clear effect |
| Use per-image/per-class metrics instead of dataset-aggregated mIoU | Semantic segmentation evaluation · Semantic segmentation datasets | bias toward large objects | Helped |
| Evaluate performance using only average metrics instead of corresponding worst-case metrics | Semantic segmentation evaluation · PASCAL Context | mIoU^C^1: mIoU^C ranges from 46.00% to 60.17%; all models have 0% mIoU^C^1 | Mixed |
Closest papers by meaning
- SUM Parts: Benchmarking Part-Level Semantic Segmentation of Urban Meshes (CVPR 2025 · similarity 0.87)
- Hierarchy-Agnostic Unsupervised Segmentation: Parsing Semantic Image Structure (NeurIPS 2024 · similarity 0.87)
- MaSS13K: A Matting-level Semantic Segmentation Benchmark (CVPR 2025 · similarity 0.86)
- RankSEG-RMA: An Efficient Segmentation Algorithm via Reciprocal Moment Approximation (NeurIPS 2025 · similarity 0.86)
- BFANet: Revisiting 3D Semantic Segmentation with Boundary Feature Analysis (CVPR 2025 · similarity 0.86)
- Uncertainty-aware Fine-tuning of Segmentation Foundation Models (NeurIPS 2024 · 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.