Submitted by John Lambert.

Submission data

Full nameMSeg: A Composite Dataset for Multi-domain Semantic Segmentation
DescriptionWe 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 titleMSeg: A Composite Dataset for Multi-domain Semantic Segmentation
Publication authorsJohn Lambert*, Zhuang Liu*, Ozan Sener, James Hays, Vladlen Koltun
Publication venueCVPR 2020
Publication URLhttp://vladlen.info/papers/MSeg.pdf
Input Data TypesUses Color        Uses 2D
Programming language(s)Python
HardwareQuadro RTX 6000, 24 GB RAM
Websitehttps://github.com/mseg-dataset/mseg-semantic
Source code or download URLhttps://github.com/mseg-dataset/mseg-semantic
Submission creation date27 Jul, 2020
Last edited28 Jul, 2020

2D semantic label results

Infoavg ioubathtubbedbookshelfcabinetchaircountercurtaindeskdoorfloorotherfurniturepicturerefrigeratorshower curtainsinksofatabletoiletwallwindow
permissive0.4850.5050.7090.0920.4270.2410.4110.6540.3850.4570.8610.0530.2790.5030.4810.6450.6260.3650.7480.7250.529