Publications

Jupyter Notebook is not only a popular tool for publishing data science results, but canalso be used for the interactive explanation of teaching content as well as the supervised work onexercises. In order to give students feedback on their solutions, it is necessary to check and evaluatethe submitted work. To exploit the possibilities of remote learning as well as to reduce the workneeded to evaluate submissions, we present a flexible and efficient framework. It enables automatedchecking of notebooks for completeness and syntactic correctness as well as fine-grained evaluationof submitted tasks. The framework comes with a high level of parallelization, isolation and a shortand efficient API.

We propose SAMAC, a Device-to-Device (D2D) communication scheme that can exploit Low Earth Orbit (LEO) satellites as temporarily-available infrastructure with intermediate round-trip-times for supporting D2D medium access of highly mobile ground nodes such as vehicles. We demonstrate how such a scheme can be kept backwards-compatible to, e.g., IEEE 802.11p and we demonstrate analytically and in computer simulations that, compared to unassisted D2D communication or relaying, in situations where a LEO satellite is available it can improve performance in terms of all of throughput, latency bounds, and reliability.

Three-dimensional scanning technology has been traditionally used in the medical and engineering industries, but these scanners can be expensive or limited in their capabilities. This research aimed to develop low-cost 3D scanning using rotation and immersion in a water-based fluid. This technique uses a reconstruction approach similar to CT scanners but with significantly less instrumentation and cost than traditional CT scanners or other optical scanning techniques. The setup consisted of a container filled with a mixture of water and Xanthan gum. The object to be scanned was submerged at various rotation angles. A stepper motor slide with a needle was used to measure the fluid level increment as the object being scanned was submerged into the container. The results showed that the 3D scanning using immersion in a water-based fluid was feasible and could be adapted to a wide range of object sizes. The technique produced reconstructed images of objects with gaps or irregularly shaped openings in a low-cost fashion. A 3D printed model with a width of 30.7200 ± 0.2388 mm and height of 31.6800 ± 0.3445 mm was compared to its scan to evaluate the precision of the technique. Its width/height ratio (0.9697 ± 0.0084) overlaps the margin of error of the width/height ratio of the reconstructed image (0.9649 ± 0.0191), showing statistical similarities. The signal-to-noise ratio was calculated at around 6 dB. Suggestions for future work are made to improve the parameters of this promising, low-cost technique.

Today’s systems are capable of storing large amounts of data in main memory. In-memoryDBMSs can benefit particularly from this development. However, the processing of the data fromthe main memory necessarily has to run via the CPU. This creates a bottleneck, which affects thepossible performance of the DBMS. The Processing-In-Memory (PIM) technology is a paradigm toovercome this problem, which was not available in commercial systems for a long time. However, withthe availability of UPMEM, a commercial system is finally available that provides PIM technologyin hardware. In this work, the main focus was on the optimization of the table scan, a fundamental,and memory-bound operation. Here a possible approach is shown, which can be used to optimizethis operation by using PIM. This method was then tested for parallelism and execution time inbenchmarks with different table sizes and compared to the usual table scan. The result is a table scanthat outperforms the scan on the usual CPU significantly.

Humans are able to learn to recognize new objects even from a few examples. In contrast, training deep-learning-based object detectors requires huge amounts of annotated data. To avoid the need to acquire and annotate these huge amounts of data, few-shot object detection (FSOD) aims to learn from few object instances of new categories in the target domain. In this survey, we provide an overview of the state of the art in FSOD. We categorize approaches according to their training scheme and architectural layout. For each type of approach, we describe the general realization as well as concepts to improve the performance on novel categories. Whenever appropriate, we give short takeaways regarding these concepts in order to highlight the best ideas. Eventually, we introduce commonly used datasets and their evaluation protocols and analyze the reported benchmark results. As a result, we emphasize common challenges in evaluation and identify the most promising current trends in this emerging field of FSOD.