Research map / Detection and segmentation / Open-set and few-shot detection
Reliability in Semantic Segmentation: Can We Use Synthetic Data?
Thibaut Loiseau, Tuan-Hung Vu, Mickael Chen, Patrick Pérez, Matthieu Cord
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 |
|---|---|---|---|
| evaluate segmenters on synthetic OOD images instead of the Cityscapes validation set to predict real shift robustness | 40 pretrained segmenters · ACDC night/snow | Pearson correlation with real mIoU | Improved |
| evaluate on synthetic data in small-gap domains (IDD, ACDC fog) | 40 pretrained segmenters · IDD, ACDC fog | Pearson correlation | No clear effect |
| use about 500 synthetic images | synthetic robustness benchmark · robustness correlation | Pearson correlation stability | Improved |
| swap SD1.5 generator for SDXL (ablation) | synthetic benchmark generator · Fog and India | Pearson correlation: 0.77 -> 0.89 (fog); 0.71 -> 0.93 (india) | Improved |
| swap SD1.5 generator for SDXL (ablation) | synthetic benchmark generator · Night and Snow | Pearson correlation: 0.85 -> 0.84 (night); 0.85 -> 0.82 (snow) | No clear effect |
| recalibrate with temperature scaling on synthetic shift data | 40 pretrained segmenters · ACDC shifts and IDD | ECE improvement success rate: 72.5% on IDD, over 90% on ACDC | Improved |
| calibrate with synthetic data instead of Cityscapes data | segmenters · shifted domains | ECE improvement | Improved |
| assess OOD detection with curated vs all inpainted images (ablation) | synthetic OOD object benchmark · SMIYC RoadAnomaly21 | correlation with real scores: entropy PCC 0.94 for both; curated around 0.8 on anomaly metrics | Mixed |
| train RbA with synthetic inpainted OOD objects | RbA Swin-B · SMIYC RoadAnomaly21 | AUROC: 95.6 -> 97.3 | Improved |
| curate inpainted images for OOD training (ablation) | RbA · SMIYC RoadAnomaly21 | AUROC: 97.2 (curated) vs 97.3 (all) | No clear effect |
| use synthetic OOD training data instead of copy-pasted COCO objects | RbA Swin-L · SMIYC RoadAnomaly21 | AUROC: 98.1 (ours all) vs 98.2 (COCO) | No clear effect |
| use per-class vs standard temperature scaling | calibration · synthetic calibration | ECE improvement | Mixed |
Closest papers by meaning
- Reliability in Semantic Segmentation: Are We on the Right Track? (CVPR 2023 · similarity 0.90)
- What Makes Synthetic Data Effective in Image Segmentation (ICML 2026 · similarity 0.89)
- Effective Use of Synthetic Data for Urban Scene Semantic Segmentation (ECCV 2018 · similarity 0.89)
- A Training-free Synthetic Data Selection Method for Semantic Segmentation (AAAI 2025 · similarity 0.87)
- Generalize or Detect? Towards Robust Semantic Segmentation Under Multiple Distribution Shifts (NeurIPS 2024 · similarity 0.87)
- Using Diffusion Priors for Video Amodal Segmentation (CVPR 2025 · similarity 0.86)
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.
