Fabrizio J. Piva
Eindhoven University of Technology
Mobile Perception Systems Lab (MPS)
EDL P16-25 P3: DL for Video Analysis and Surveillance (VAS)
Unsupervised Domain Adaptation for Semantic Segmentation
When training deep learning models to perform a specific task (for instance, semantic segmentation to detect objects per pixel), we usually have a labeled dataset consisting of input images along with pixel-wise label maps indicating the class that a particular pixel belongs to. Nowadays, thanks to the advancement of deep learning models, it is possible to achieve highly accurate results on any particular dataset when enough labeled examples are provided. However, if we wish to use this trained model on new real-world data from a different domain, the performance is likely to drop significantly due to the domain gap between the source training data and the target real-world data.
Considering the burden that involves manually labeling data and the convenience of being able to collect virtually an infinite of unlabeled examples, the goal of Unsupervised Domain Adaptation (UDA) is to build a model able to learn from both labeled data from a source domain and unlabeled data from a target domain. To achieve this, several techniques are applied to either learn domain-invariant features across both domains or to transfer the learned knowledge from the source to the target domain.
The resulting UDA model will ultimately perform significantly better than the one trained only on the source labeled dataset, making it production-friendly for real-world applications.