| Full name | MSeg: A Composite Dataset for Multi-domain Semantic Segmentation |
| Description | We present MSeg, a composite dataset that unifies semantic segmentation datasets from different domains. A naive merge of the constituent datasets yields poor performance due to inconsistent taxonomies and annotation practices. We reconcile the taxonomies and bring the pixel-level annotations into alignment by relabeling more than 220,000 object masks in more than 80,000 images, requiring more than 1.34 years of collective annotator effort. The resulting composite dataset enables training a single semantic segmentation model that functions effectively across domains and generalizes to datasets that were not seen during training. We adopt zero-shot cross-dataset transfer as a bench- mark to systematically evaluate a model’s robustness and show that MSeg training yields substantially more robust models in comparison to training on individual datasets or naive mixing of datasets without the presented contributions. |
| Publication title | MSeg: A Composite Dataset for Multi-domain Semantic Segmentation |
| Publication authors | John Lambert*, Zhuang Liu*, Ozan Sener, James Hays, Vladlen Koltun |
| Publication venue | CVPR 2020 |
| Publication URL | http://vladlen.info/papers/MSeg.pdf |
| Input Data Types | Uses Color Uses 2D |
| Programming language(s) | Python |
| Hardware | Quadro RTX 6000, 24 GB RAM |
| Website | https://github.com/mseg-dataset/mseg-semantic |
| Source code or download URL | https://github.com/mseg-dataset/mseg-semantic |
| Submission creation date | 27 Jul, 2020 |
| Last edited | 28 Jul, 2020 |
