Research map / Label-efficient and robust learning / Domain adaptation and generalization
Effective Use of Synthetic Data for Urban Scene Semantic Segmentation
Fatemeh Sadat Saleh, Mohammad Sadegh Aliakbarian, Mathieu Salzmann, Lars Petersson, Jose M. Alvarez
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 |
|---|---|---|---|
| add detection-based (Mask R-CNN) handling of foreground classes to DeepLab trained on GTA5+VEIS | DeepLab trained on synthetic data · Cityscapes val | mIoU: 32.8 -> 38.0 | Improved |
| add VEIS data to GTA5 training in a plain semantic segmentation network | DeepLab · Cityscapes val | mIoU: 31.3 -> 32.8 | Improved |
| add pseudo-labels on unlabeled real images to the plain GTA5+VEIS baseline | GTA5+VEIS DeepLab · Cityscapes val | mIoU: 32.8 -> 33.3 | Improved |
| add pseudo-labels from the detection-based method on unlabeled real images | Ours · Cityscapes val | mIoU: 38.0 -> 42.5 | Improved |
| add pseudo-labels from the detection-based method on unlabeled real images | Ours · CamVid test | mIoU: 47.6 -> 48.8 | Improved |
| add detection-based foreground handling with VEIS to GTA5 baseline | GTA5 DeepLab · CamVid test | mIoU: 44.6 -> 47.6 | Improved |
| add VEIS data to GTA5 in a plain segmentation network | GTA5 DeepLab · CamVid test | mIoU: 44.6 -> 44.4 | No clear effect |
| replace GTA5 by VEIS as training set for a plain segmentation network | DeepLab · Cityscapes val | mIoU: 31.3 -> 24.4 | Hurt |
| apply the detection-based approach when training on Cityscapes real labeled data | fully supervised DeepLab · Cityscapes val | mIoU: 56.2 -> 60.0 | Improved |
| use texture instead of silhouette as classifier input for real vs synthetic discrimination | binary VGG-16 classifier · VEIS vs Cityscapes objects | accuracy: 70.0% -> 95.1% | Improved |
| train object classifier on textured silhouettes instead of plain silhouettes | multi-class VGG-16 classifier · Cityscapes real objects test | accuracy gap real vs synthetic: 81.0 vs 89.2 (silhouette); 83.7 vs 94.2 (textured) | Hurt |
Closest papers by meaning
- ROAD: Reality Oriented Adaptation for Semantic Segmentation of Urban Scenes (CVPR 2018 · similarity 0.93)
- Learning Semantic Segmentation From Synthetic Data: A Geometrically Guided Input-Output Adaptation Approach (CVPR 2019 · similarity 0.92)
- The SYNTHIA Dataset: A Large Collection of Synthetic Images for Semantic Segmentation of Urban Scenes (CVPR 2016 · similarity 0.92)
- A Curriculum Domain Adaptation Approach to the Semantic Segmentation of Urban Scenes (TPAMI 2019 · similarity 0.92)
- Not All Areas Are Equal: Transfer Learning for Semantic Segmentation via Hierarchical Region Selection (CVPR 2019 · similarity 0.91)
- A Training-free Synthetic Data Selection Method for Semantic Segmentation (AAAI 2025 · similarity 0.88)
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.
