This work presents a two-staged, unsupervised domain adaptation process for semantic segmentation models by combining a self-training and self-supervision strategy. Self-training (i. e., training a model on self-inferred pseudolabels) yields competitive results for domain adaptation in recent research. However, self-training depends on high-quality pseudo-labels. On the other hand, self-supervision trains the model on a surrogate task and improves its performance on the target domain without further prerequisites.Therefore, our approach improves the model's performance on the target domain with a novel surrogate task. To that, we continuously determine class centroids of the feature representations in the network's pre-logit layer on the source domain. Our surrogate task clusters the pre-logit feature representations on the target domain regarding these class centroids during both training stages. After the first stage, the resulting model delivers improved pseudo-labels for the additional self-training in the second stage. We evaluate our method on two different domain adaptions, a real-world domain change from Cityscapes to the Berkeley Deep Drive dataset and a synthetic to real-world domain change from GTA5 to the Cityscapes dataset. For the real-world domain change, the evaluation shows a significant improvement of the model from 46% mIoU to 54% mIoU on the target domain. For the synthetic to real-world domain change, we achieve an improvement from 38.8% to 46.42% on the real-world target domain.

Recent years have seen major advances in Artificial Intelligence (AI) methods for environment perception in intelligent transportation systems. Although most of them have been achieved in the automotive sector there is a similar demand in the railway domain. This paper investigates Deep Neural Network (DNN) based environment perception using vehicle-borne camera images from the rail domain. Specifically, railway switch detection and classification are addressed as a relevant example for a DNN application with potential use for landmark positioning, environment perception, and condition monitoring. The lack of large training data sets in the railway sector (in contrast to the automotive domain) is compensated by an appropriate DNN architecture, an anchor box ratio optimization scheme, and transfer learning. The presented experimental results advocate for the adopted approach.

The performance of a segmentation network optimized on data from a specific type of OCT sensor will decrease when applied to data from a different sensor. In this work, we deal with the research question of adapting models to data from an unlabeled new sensor with new properties in an unsupervised way. This challenge is known as unsupervised domain adaptation and can alleviate the need for costly manual annotation by radiologists. We show that one can strongly improve a model's result that was trained in a supervised way on the source OCT sensor domain on the target sensor domain. We do this by aligning the source and target domain distributions in the feature space through a semantic clustering method. Apart from the unsupervised domain adaptation, we improved even the supervised training compared to the results in the RETOUCH challenge by employing a sophisticated training strategy. The RETOUCH challenge contains three different types of OCT scanners and provides annotations for the task of disease-related fluid classes.

Roundabouts are well-known for their ability to improve upon traffic safety, especially for motorized traffic. An in-depth analysis on this topic is known from previous work. It was found that different types of roundabouts have different levels of safety. The work at hand is a replication study for a previous study in this regard. It uses a mix of traditional and a Machine Learning (ML)-based approach, expands on the previous results and replicates some of the previous findings. This was possible especially by using a factor of 10 more roundabouts in the analysis, with considerably less manual intervention. Furthermore, this study could also draw some additional conclusions regarding the safety of bicyclists, which were not included in the original study. Finally, by using cross-validation techniques, a kind of minimal model could be established that needs fewer factors and achieves better prediction quality than straightforward glm models.

Active learning automatically selects samples for annotation from a data pool to achieve maximum performance with minimum annotation cost. This is particularly critical for semantic segmentation, where annotations are costly. In this work, we show in the context of semantic segmentation that the data distribution is decisive for the performance of the various active learning objectives proposed in the literature. Particularly, redundancy in the data, as it appears in most driving scenarios and video datasets, plays a large role. We demonstrate that the integration of semisupervised learning with active learning can improve performance when the two objectives are aligned. Our experimental study shows that current active learning benchmarks for segmentation in driving scenarios are not realistic since they operate on data that is already curated for maximum diversity. Accordingly, we propose a more realistic evaluation scheme in which the value of active learning becomes clearly visible, both by itself and in combination with semisupervised learning.