Deformable Faster-RCNN for Lesion Detection in CT Images (Master's thesis)

Speaker: Fabio Hellmann

Time: 10:00, Friday 11th December 2020

Location: Online



An Analysis of Bias and Fairness in Affective Computing Orientated Machine Learning (Master's thesis)

Speaker: Niklas Schröter

Time: 10:00, Friday 11th December 2020

Location: Online



Webcam-basierte Herzfrequenzmessung mithilfe neuronaler Netze (Bachelor's thesis)

Speaker: Jonas Mayr

Time: 10:00, Friday 11th December 2020

Location: Online



Mobile Audio Processing and Event Detection Utilising Energy Efficient Deep Neural Networks (Master's thesis)

Speaker: Tobias Hübner

Time: 16:30, Monday 19th October 2020

Location: Online



An investigation of sequence modelling strategies for speech-based depression detection (Bachelor's thesis)

Speaker: Carmen Berndt 

Time: 10:00, Wednesday 17th June 2020

Location: Online




University Library Services - Publication lists & self-archiving options


The University Library Service now offers two new services centered around the better visibility of your research publications. During this brief, informal presentation Sonja Härkönen will provide an overview of the services available and of how they can be accessed. The focus will be on the display of publication lists on the new website (using a plugin for DjangoCMS) and on the self-archiving options available (i.e. the option to make the full text of your publications freely available online).


Speaker: Sonja Härkönen, Librarian, Augsburg University

Time: 13:45, Friday 21st February 2020

Location: Room 004, Eichleitnerstr. 30, F1




Generative–Temporal based Model for Affect Estimations


The temporal information, which describes the sequential dependency across the frames naturally exists in the video-based datasets. In several machine vision tasks, including facial based analysis, these inherent features have been shown to be beneficial to model training in improving their accuracy. In affective computing, however, deep investigations of these methods are yet to be undergone given its unique challenges. One instance is that it demands more sophisticated understanding of complex interactions between humans besides of the accuracy of the model. To this end, the current task will be first to analyse of the impact of temporal modeling through generative modeling in current affect estimations problem. One such approach is to use deep auto-encoder based structure to model inherent features from each modality. Then, recurrent based deep learning models, such as RNN and LSTMs, can be used to capture relevant temporal information and final estimations through curriculum learning. Further, utilizing these findings, necessary machine learning-based techniques will be integrated to achieve the main aim of state of the art accuracy on the available benchmarks.



Decky Aspandi is currently a Ph.D candidate in the Department of Information and Communication Technologies at the Universitat Pompeu Fabra (UPF), Barcelona. He is supervised by Xavier Binefa and funded by the Predoctoral Fellowship of the UPF. In 2014 he received his Master Degree in Computer Engineering from King Mongkuts University of Technology Thonburi, Thailand. In 2012 he received his Bachelor in Computer science from the University of Mulawarman, Indonesia. He has taught several undergraduate courses including Computer Organization and Deep Learning. He is currently working on his thesis, which focuses on the understanding of facial analysis using deep spatio-temporal neural network models.


Speaker: Decky Aspandi

Time: 10:00, Thursday 16th January 2020

Location: Room 004, Eichleitnerstr. 30, F1

Contribution Analysis for Exterior Noise Development


The exterior noise is one of the key components in the automotive NVH development process. While strict regulations have to be met, the acoustic character is of major importance for the OEMs. The simulated pass-by is a widely accepted alternative to the real test track, as it is more convenient and independent from environmental conditions. Modifications and additional measurement tasks can be added easily in order to minimize possible sources of disturbances. With the so-called contribution analysis, a ranking of the most common sources, such as intake, engine, exhaust, and tires can be easily performed in addition to the standard measurements. This indicates at which positions modifications might make sense. The process to determine these contributions is rather tiresome, as all possible sources and transfer paths have to be measured with microphones and accelerometers, in order to decompose the response at the AA line. Hence, a machine learning based approach is desired, In order to guide engineers to critical sound and vibration sources way faster. The task is now to train a system, which is able to decompose the exterior noise into tires, engine, exhaust, gear box and powertrain. Further, an easy transfer to comfort noise measurements within the vehicle is aimed at.



Dejan Arsić received his diploma in 2004, followed by his doctoral degree in 2010 for his studies in audiovisual signal processing with focus on object tracking and behavioral analysis at the faculty of electrical engineering at Technische Universität München (TUM). He currently applies his profound knowledge in signal analysis as key account manager in the automotive industry. Here he establishes elaborate solutions for various problems in the field of NVH, with a main focus on TPA methods and workflow oriented data acquisition and analysis. Dr. Arsić has (co-)authored more than 60 publications in books, journals and conference proceedings, and is reviewer for a range of leading scientific journals.


Speaker: Dr. Dejan Arsić, Müller-BBM VibroAkustik

Time: 10:00, Wednesday, 8th January 2020

Location: Room 004, Eichleitnerstr. 30, F1