Browsing by Author "Laakso, Joosua"

Semantic segmentation is a computer vision problem of partitioning an image based on what type of an object each part represents, with pixel-level precision. Producing labeled datasets to train deep learning models for semantic segmentation can be laborious due to the demand for pixel-level precision. On the other hand, a deep learning model trained on one dataset might have inferior performance when applied on another dataset, depending on how different those datasets are. Unsupervised domain adaptation attempts to narrow this performance gap by adapting the model to the other dataset, even if ground-truth labels for that dataset are not available. In this work, we review some of the pre-existing methods for unsupervised domain adaptation in semantic segmentation. We then present our own efforts to develop novel methods for the problem. Those include a new type of loss function for unsupervised output shaping, unsupervised training of the model backbone based on the feature statistics and a method for unsupervised adaptation of the model backbone using an auxiliary network that attempts to mimic the gradients of supervised training. We present empirical results of the performance of these methods. We additionally present our findings on the effects of changes in the statistics of the batch normalization layers on domain adaptation performance.