News Archive

Aktuelles Archiv

Abgabe-Präsentation der Masterarbeit von Simone Wälde

Datum: Dienstag, 07.07.2020

Uhrzeit: 14:00 Uhr

Ort: Virtuell: MS Teams Meeting*

Titel: Geometry Classification through Feature Extraction and Pattern Recognition in 3D Space

Zusammenfassung:

In dieser Masterarbeit wird der Versuch unternommen, ähnliche Wappendarstellungen auf 3D-Modellen von Scherben abzugleichen.

Teil eines initialen Workflows ist die Reliefextraktion, für die ein Ansatz von Zatzarinni et al. verwendet wird.

Um Informationen der Objektoberfläche zu extrahieren, wird eine Local Binary Pattern-Variante von Thompson et al. implementiert.

Die resultierenden Merkmalsdeskriptoren werden dann unter Verwendung einer Abstandsmetrik verglichen.

Am Ende führt der vorgeschlagene Ansatz nicht zu guten Ergebnissen, aber die aufgetretenen Herausforderungen sind dokumentiert und Vorschläge für zukünftige Lösungen werden diskutiert.

Abgabe-Präsentation der Masterarbeit von Hasan Kutlu

Datum: Dienstag, 07.07.2020

Uhrzeit: 11:00 Uhr

Ort: Virtuell: MS Teams Meeting*

Titel: Fully Automatic Mechanical Scan Range Extension of a Lens-Shifted Structured Light System

Zusammenfassung:

Der auf strukturiertem Licht basierende MesoScannerV2 erreicht eine erweiterte Tiefen- und laterale Auflösung, aufgrund seiner Besonderheit der Erweiterung aktueller Streifenmuster-Ansätze um eine mechanische, linsenverschobene Oberflächenkodierungsmethode.

Aufgrund von optischen Herausforderungen in der Datenerfassung durch die hohe Auflösung und möglicher Unsicherheiten der numerischen Algorithmen entsteht Rauschen, das die digitalisierten 3D-Modelle direkt beeinflusst.

Ziel dieser Arbeit war es daher, das erzeugte Rauschen zu reduzieren, um sauberere 3D-Modelle zu erhalten.

Die Integration eines Verfahrens zur Automatisierung des Scanprozesses vieler Objekte gleichzeitig ist ein weiteres Thema, das in dieser Arbeit behandelt wird.

Wir zeigen, dass Methoden gefunden werden können, um das erzeugte Rauschen deutlich zu reduzieren, und liefern eine entsprechende Auswertung.

Weiterhin werden mögliche Lösungen zur Automatisierung des Scanprozesses vorgestellt.

* Die Nutzungsbedingungen und *Datenschutzinformationen* zur Teilnahme an „Teams“-Meetings finden Sie unter:

https://fh-igd.de/FO365_Nutzungsbedingungen

Habilitationsvortrag

Zeit: 19.03.2020, 12:00

Ort: TU Darmstadt, FB Informatik, Hochschulstraße 10, 64289 Darmstadt, Piloty-Gebäude S2|02, Raum C110

Referent: Dr.rer.nat. Stefan Guthe

Titel: “The Jump Point Search Algorithm and its Applications” (Vortrag)
“Applied Perception for Real-time Computer Graphics” (Habilitationsarbeit)

Link: https://www.informatik.tu-darmstadt.de/fb20/aktuelles_fb20/veranstaltungen_fb20/veranstaltungen_fb20_details_187392.de.jsp

Dissertationsvortrag

Zeit: 29.04.2020, 15:00

Ort: Raum 074 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Johannes Fauser

Titel: “Preoperative Surgical Planning
Toward an Automatic Pipeline for Segmentation and Nonlinear Trajectory Planning in Robot-Assisted Interventions

Abschlussarbeiten

Zeit: 03.03.2020, 13:00

Ort: Raum 220 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Speaker: Andreas Giebel (Betreuer: Tim Grasser)

Titel: ”GPU-beschleunigte Finite Elemente Modalanalyse mit Projektionsansatz“ (Bachelorarbeit)

Abstract: Diese Arbeit befasst sich mit der numerischen Bestimmung der Eigenschwingungsparameter eines schwingungsfähigen Systems. Mithilfe von Finite Elemente Diskretisierungen kann ein generalisiertes Eigenwertproblem für dieses System formuliert werden, welches mit iterativen, approximierenden Eigenwertverfahren wie dem generalisierten Lanczos-Verfahren gelöst werden kann. Dieses Verfahren kann effizient parallelisiert werden, durch Parallelisierung der zugrundeliegenden Operationen der linearen Algebra. Besonders zeitintensiv ist dabei das Lösen linearer Gleichungssysteme, was für Massivparallele Hardware, wie Grafikkarten, mit dem modifizierten PCG-Verfahren speichereffizient und skalierbar implementiert werden kann. Durch Nutzung eines Projektionsansatz können Randbedingungen implizit angewandt werden ohne Matrixmanipulationen durchzuführen. Die GPU-Version erzielte einen maximalen Speedup von ca. 24, 71 gegenüber den CPU-Implementierungen, bei einem Speicherverbrauch von nur ca. einem Siebtel. Schließlich werden die Ergebnisse der entwickelten Modalanalyse mit analytischen Ergebnissen und den Simulationsergebnissen von Ansys anhand eines Balkens verglichen.

Zeit: 17.02.2020, 14:00

Ort: Raum 011 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Thomas Pöllabauer (Betreuer: Pavel Rojtberg)

Titel: ”STYLE: Style Transfer for Synthetic Training of a YoLo6D Pose Estimator“ (Masterarbeit)

Abstract: This work deals with training of a DNN-based object pose estimator in three scenarios: First, a small amount of real-world images of the objects of interest is available, second, no images are available, but object specific texture is given, and third, no images and no textures are available. Instead of copying successful randomization techniques, these three problems are tackled mainly with domain adaptation techniques.

The main proposition is the adaptation of general-purpose, widely-available, pixel-level style transfer to directly tackle the differences in features found in images from different do- mains. To that end several approaches are introduced and tested, corresponding to the three different scenarios. It is demonstrated that in scenario one and two, conventional conditional GANs can drastically reduce the domain gap, thereby improving performance by a large mar- gin when compared to non-photo-realistic renderings. More importantly: ready-to-use style transfer solutions improve performance significantly when compared to a model trained with the same degree of randomization, even when there is no real-world data of the target objects available (scenario three), thereby reducing the reliance on domain randomization.

Gastvorlesung

Zeit: 14.02.2020, 15:30

Ort: Raum 072 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Adam Blitz
Near Eastern Studies and Judaic Studies scholar, London

Titel: ”Digital Apamea: Digital Reconstruction of Ancient 4th Century Synagogue Mosaic Floor“

Abstract:

Digital Apamea is both an actual and virtual exhibition in which digital technology is employed to reconstruct the ancient 4th Century synagogue mosaic floor at Apamea on Orontes in Syria.

The synagogue was built in 392 CE (AD). Shortly thereafter it was intentionally destroyed at a time when Roman tolerance yielded to Byzantine intolerance and anti-Semitism per the edicts issued by the emperor Theodosius the Second. Two subsequent basilicae were built over the site of the synagogue. These churches incorporated the former synagogue foundations into the latter structures; the synagogue’s mosaic floor (with Greek donor inscriptions) being buried intact in an act of Damnatio Memoriae. Apamea’s synagogue remained unknown to the world until 1934. Thereafter the mosaics were detached from context and divided between various stakeholders. The excavation reports were destroyed in a fire in Brussels in the late 1940’s. Yet today we can recreate the floor on account of digital technology.

Digital Apamea tells the story of multi-culturalism in Roman Syria, hatred under Byzantine rule, contested space, the archaeology and architecture of Late Antique synagogues and churches, the history of mosaic execution, digital technology and its use for archaeologists (notably as tool to monitor and capture images of illegal excavation at site).

Digital Apamea also explores two artistic legacies: the on-going tradition of Romano-Byzantine mosaic design in Ravenna and also an initiative by the Latin Patriarchate of Jerusalem in Madaba, Jordan. In the revered city of Madaba, the Patriarchate together with CARITAS runs the “Living Mosaic Project” to train Iraqi refugees from Mosul as mosaicists.

Digital Apamea at CHNT Visual Heritage 23 shares a virtual gallery with visitors and allows the participants to experience what would otherwise be the physical exhibition (2018: Vienna; 2019, Oxford (UK), Barcelona, Göttingen, Mostra Mercato Bienno, Italy). It also includes video, presentation and Virtual Reality (VR) experience for all ages.

Digital Apamea©2018 Finalist: International Symposium on Cultural Heritage Conservation and Digitalization, Tsinghua University, Beijing 2018.

For further enquiries please contact info@digitalapamea.org or adam.blitz@ronininstitute.org.

www.digitalapamea.org

Bio:

Adam Blitz is a Near Eastern Studies and Judaic Studies scholar and former Fulbright recipient (Germany). He is a Fellow of the Royal Anthropological Institute, London and also The Ronin Institute, New Jersey, USA. He resides in London. He holds degrees from the University of Edinburgh, University College London, University of Michigan, Humboldt Universitaet zu Berlin and Birkbeck College, London. The majority of his research examines cultural heritage, archaeology/ anthropology, the Ancient Near East and the Middle East with particular focus on Syria – notably its ancient and modern synagogues. He has written over 30 articles for various publications to include: The Times of Israel, The Jerusalem Post, Haaretz, Catholic Herald, Aeon Magazine, Times of India, Anthropology Today, 9 Muses News etc. He is a committee member of ICOMOS ICORP-UK (International Scientific Committee on Risk Preparedness) and an active member of English PEN. He is also an accredited journalist with the European News Agency (ID 13412)

He can be contacted at adam.blitz@ronininstitute.org / adam.blitz@columnist.com.

Abschlussarbeiten

Zeit: 04.02.2020, 14:00

Ort: Raum 242 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Tim Alexander Bergmann (Betreuer: Matthias Noll)

Titel: ”AR-Visualisierung von Echtzeitbildgebung für ultraschallgestützte Leberbiopsien“ (Masterarbeit)

Abstract: In dieser Arbeit wird ein Augmented Reality-System zur Anzeige von Ultraschallbildern direkt am Patienten vorgestellt. Die Überlagerung wird mit der Hilfe von optisch durchsichtigen Head-mounted Displays durchgeführt. Die lagegerichtete Darstellung der Ultraschallbilder basiert auf einem externen optischen Trackingsystem, dem NDI Polaris Vicra. Um die korrekte Überlagerung zu gewährleisten, wird das Sichtfeld eines Trägers mittels angepasster Single Point Active Alignment Method bestimmt. Die Lage der Ultraschallbilder relativ zu den Tracking-Markierungen der Ultra-schallsonde wird mit einer angepassten Pivot-Kalibrierung ermittelt. Zum objektiven Testen des Systems wurde ein Träger-Dummy verwendet, der das Sehen eines Trägers durch Kameras simuliert. Die Lage von Tracking-Markierungen im Sichtfeld des Träger-Dummies konnte mit einem RMSE von 1,1480 mm bestimmt werden. Bei den Tests der Überlagerung der Ultraschallbilder über den darin repräsentierten Strukturen erreicht das System einen Dice-Koeffizienten von 88,33 %. Zur besseren Skalierung der Berechnungsdauer mit der Anzahl der verwendeten Geräte wurden Matrixoperatoren für die verwendeten Transformationsmatrizen optimiert. Die Berechnungen werden im Schnitt mehr als dreimal so schnell durchgeführt wie die allgemeine Implementierung der Operatoren. Das System versetzt behandelnde Ärzte in die Lage, Ultraschallbilder lagegerichtet über den darin repräsentierten Strukturen zu betrachten. Die Anzeige der Bilder auf einem externen Monitor wird dadurch überflüssig.

Keywords:

affine transformation, augmented reality(AR), calibration, computer assisted surgery, field of view simulation, head mounted displays, image guided therapy, infrared tracking systems, intraoperative navigation, liver, marker based tracking, matrix operations, medical imaging, MedVis, minimally invasive surgery, OpenGL, optical tracking, ultrasound

Zeit: 20.01.2020, 14:00

Ort: Raum 048 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Kai Alexander Neumann (Betreuer: Matevz Domajnko)

Titel: ”Adaptive Camera View Clustering for Fast Incremental Image-based 3D Reconstruction“ (Bachelorarbeit)

Abstract: In this thesis, we present an evaluation of different state-of-the-art photogrammetry solutions to identify which of them is most capable of providing feedback that predicts the quality of the final 3D reconstruction during acquisition. For this, we focused on the open-source incremental reconstruction solutions COLMAP, Alicevision Meshroom and MVE. Additionally, we included the commercial solution Agisoft Metashape to evaluate how it compares against the open-source solutions.

While we were able to identify some characteristic behaviors, the accuracy and runtime of all four reconstruction solutions vary based on the input dataset. Because of this, and the fact that all four solutions compute very similar results under the same conditions, our tests were not conclusive. Nevertheless, we chose COLMAP as the back-end for further use as it provided good results on the real dataset as well as an extensive command-line interface (CLI).

Based on these results, we introduce an iterative image-based reconstruction pipeline that uses a cluster-based acceleration structure to deliver more robust and efficient 3D reconstructions. The photogrammetry solution used for reconstruction is exchangeable. In this pipeline, images that portray common parts of an object are assigned to clusters based on their camera frustums. Each cluster can be reconstructed separately. The pipeline was implemented as a c++ module and tested on the autonomous robotic scanner CultArm3D R . For this system, we embedded the pipeline in a feedback loop with a density-based Next-Best-View (NBV) algorithm to assist during autonomous acquisition.

Zeit: 10.01.2020, 16:00

Ort: Raum 324 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Camila Gonzalez (Betreuer: Anirban Mukhopadhyay)

Titel: ”Preventing Catastrophic Forgetting in Deep Learning Classifiers“ (Masterarbeit)

Abstract: Deep neural networks suffer from the problem of catastrophic forgetting. When a model is trained sequentially with batches of data coming from different domains, it adapts too strongly to properties present on the last batch. This causes a catastrophic fall in performance for data similar to that in the initial batches of training.

Regularization-based methods are a popular way to reduce the degree of forgetting, as they have an array of desirable properties. However, they perform poorly when no information about the data origin is present at inference time. We propose a way to improve the performance of such methods which comprises introducing insularoty noise in unimportant parameters so that the model grows robust against them changing.

Additionally, we present a way to bypass the need for sourcing information. We propose using an oracle to decide which of the previously seen domains a new instance belongs to. The oracle’s prediction is then used to select the model state. In this work, we introduce three such oracles.

Two of these select the model which is most confident for the instance. The first, the cross-entropy oracle, chooses the model with least cross-entropy between the prediction and the one-hot form of the prediction. The second, the MC dropout oracle, chooses the model with lowest standard deviation between predictions resulting from performing an array of forward passes while applying dropout.

Finally, the domain identification oracle extracts information about the data distribution for each task using the training data. At inference time, it assesses which task the instance is likeliest to belong to, and applies the corresponding model.

For all of our three different datasets, at least one oracle performs better than all regularization-based methods. Furthermore, we show that the oracles can be combined with a sparsification-based approach that significantly reduces the memory requirements.

Zeit: 19.12.2019, 13:00

Ort: Raum 011 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Steven Lamarr Reynolds (Betreuer: Gerrit Voss)

Titel: ”Integrating physically based rendering (PBR) techniques into a multi-object-draw rendering framework“ (Bachelorarbeit)

Abstract: The current rendering system utilized by instant3Dhub breaks with the more traditional scheme of drawing one object with a fixed material assignment at a time . In order to utilize culling techniques (e.g. occlusion culling, small feature culling) more efficiently, objects are presorted spatially, cut, and parts of them are rearranged into different draw calls, if needed. At runtime a single draw call will therefore process multiple objects each with an individual material. All material information is collected in a global gpu shader storage array, which is activated for all draw calls at the beginning of a frame.

While this works well with untextured, Blinn-Phong shading model based engineering materials, extending it to include fully textured PBR techniques poses itself a set of challenges with respect to resource and GPU state management as the required state must contain at least all objects within a single draw call into account. This combined state must be actived before the drawcall is executed.

The tast of this thesis was to choose and analyse a PBR technique with respect to resource requirements, global as well as object local. To derive proposal(s) for new instant3Dhub rendering pipeline(s) with accordingly adjusted resource management, taking the PBR requirements into account. And to evaluate possible implementations.

Zeit: 12.12.2019, 16:50

Ort: Raum 072 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Steven Lamarr Reynolds (Betreuer: Marija Schufrin)

Titel: ”A Visualization Interface to Improve the Transparency of Collected Personal Data on the Internet“ (Masterarbeit)

Abstract: In den letzten Jahrzehnten hat das Internet ein enormes Wachstum erlebt. Die Nutzung von Online-Diensten im alltäglichen Leben für allerlei Arten von Aktivitäten steigt. In den meisten Fällen sammeln Unternehmen dabei Daten zur Bereitstellung und Verbesserung ihrer Online-Dienste. Viele Nutzer sind sich jedoch nicht über das Ausmaß und die Art der bei der Nutzung erhobenen und gespeicherten Daten bewusst.

Diese Arbeit ist der Fragestellung gewidmet, wie der durchschnittliche Internetnutzer darin unterstützt werden kann, einen Einblick in die von ihm gesammelten Daten zu erhalten und damit einen bewussteren Umgang zu entwickeln. Ziel der dabei entwickelten Anwendung ist es, die Transparenz der gesammelten Daten zu verbessern. Dazu wurde ein Überblick darüber gewonnen, wie Daten aktuell im Internet gesammelt werden. Basierend auf diesem Überblick wurde für diese Arbeit als Datengrundlage die personenbezogene Daten von Online-Diensten ausgewählt, die Nutzer mit dem Auskunftsrecht der europäischen Datenschutz-Grundverordnung (DSGVO) anfordern können. Darauf aufbauend wurden Konzepte entwickelt, wie diese Daten mit Hilfe von Informationsvisualisierung für den durchschnittlichen Internetnutzer visualisiert werden können. Das am besten geeignete Konzept wurde ausgewählt und prototypisch als Webinterface implementiert. Nutzer können von mehreren ausgewählten Online-Plattformen ihre Daten einfügen und visuell explorieren. Durch die ausgewählte Zusammensetzung von interaktiven Visualisierungsansätzen soll bei der Einschätzung der Menge, der Typen und beim Erkunden von Mustern oder Trends unterstützt werden. Eine zusätzliche Funktion erlaubt es dem Nutzer, die einzelnen Datenelemente nach der wahrgenommenen Sensibilität zu bewerten. Hierdurch soll der Nutzer bewusst dazu angeregt werden, sich mit seinen Daten auseinanderzusetzen. Die implementierte Anwendung wurde im Rahmen dieser Arbeit mit 37 echten Nutzern und deren persönlichen Daten evaluiert. Aus den Ergebnissen lässt sich ableiten, dass die Nutzung der Anwendung einen Einfluss auf die Haltung zur Privatsphäre in Online-Diensten der Teilnehmer hatte und somit einen möglichen Weg in Richtung besseres Bewusstsein für die eigene Privatsphäre im Internet darstellt.

Zeit: 12.12.2019, 16:00

Ort: Raum 072 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Salmah Ahmad (Betreuer: Marija Schufrin)

Titel: ”A Design Study on the Joy of Use in Information Visualization for Cybersecurity Analysis for Home Users“ (Masterarbeit)

Abstract:

In der heutigen Gesellschaft sind die meisten Menschen mindestens einmal täglich online, während das allgemeine Bewusstsein für Cybersicherheit noch gering ist. Vor allem für durchschnittliche Nutzer ohne viele IT-Kenntnisse kann es schwierig sein, sich mit dem Bereich der Cybersicherheit zu befassen. Diese Arbeit nutzt den Joy of Use-Ansatz, um die Selbstbeteiligung und die Freude für die Nutzer zu erhöhen. Dazu werden mehrere Methoden, die der Joy of Use nahe sind, beschrieben und Aspekte von Joy of Use erarbeitet.

Auf der Grundlage einer bestehenden Informationsvisualisierungsschnittstelle für die Erkundung des Netzwerkverkehrs für Heimnutzer untersucht diese Arbeit, welche Strategie die Joy of Use am besten steigern kann.

Um die Lücke zwischen Laien und dem Bereich der Cybersicherheit zu vermindern, werden drei Konzepte entwickelt, die auf der bereits existierenden Schnittstelle zur Visualisierung von Netzwerkdaten aufbauen. Diese Konzepte nutzen Joy of Use-Methoden, um die Freude eines Benutzers zu steigern. Eine Online-Vorstudie wird durchgeführt, um den Ziel-Nutzer zu charakterisieren. Dies hilft bei der Entwicklung der Konzepte unter Berücksichtigung der Bedürfnisse des Ziel-Nutzers. Jedes dieser Konzepte konzentriert sich auf eine Strategie zur Steigerung der Joy of Use in Bezug auf das Auffinden unsicherer Verbindungen im eigenen Netzwerkverkehr. Diese helfen dann bei der Entwicklung der Prototypen, die in einer Designstudie ausgewertet werden. Die anschließende Designstudie wird als Nutzer-Studie vor Ort durchgeführt, bei der die Teilnehmer nach der Verwendung jedes Prototypen ihren emotionalen Zustand und die wahrgenommene Joy of Use bewerten. Die Ergebnisse werden statistisch ausgewertet. Diese Arbeit schließt damit, dass die Verwendung von Joy of Use-Techniken bei der Präsentation unzugänglicher Themen den Heimanwendern helfen kann, Bewusstsein und Initiative beim Lernen über Cybersicherheit zu gewinnen.

Zeit: 09.12.2019, 15:00

Ort: Raum 324 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Sofie Hofmann (Betreuer: Anirban Mukhopadhyay)

Titel: ”Neural Networks for Error Correction in Electromagnetic Trackings“ (Bachelorarbeit)

Abstract: Electromagnetic tracking systems enable real-time tracking of medical equipment without line-of-sight. Thus, they are a key technology for navigation in minimally invasive surgery. Examples for applications are bronchoscopy [1] and endovascular surgery [2, 3, 4]. Submillimetric accuracy while tracking is required especially in ENT (ear, nose, throat) surgery [5]. However, the accuracy of the tracking system is very sensitive to the environment and error causing objects like ferromagnetic or other conductive materials are hard to avoid in clinical scenarios. For this reason, a main task of electromagnetic tracking applications is error compensation in order to ensure that especially risk structures are not damaged while navigating surgical instruments.

This work aims to correct positional errors caused by electromagnetic tracking systems with neural networks. A feed-forward neural network is developed that reduces errors in different environments. For this purpose, five different data sets are acquired during this work: One data set in laboratory environment and four data sets in vicinity to an X-ray machine which has a cast steel enclosure. For each data set, the X-ray machine is moved to another position. In total, each data set consists of around 700 tracked positions. These data sets are acquired with a LEGO phantom and the 3D Guidance tracking system from Ascension Technology. Wrong and inaccurate corrections are very risky for the patient, that is why an additional feedforward neural network is developed which makes use of Monte Carlo dropout [6] to approximate uncertainty of the correction algorithm. The developed neural networks are examined in different experiments: The number of training samples is varied as well as their spatial distribution in order to evaluate their robustness towards less dense training sets. In addition, extrapolation experiments are performed and transfer learning is examined – the correction of samples which originate from another data set as the network is trained with. These experiments are also basis of the final comparison between neural network and polynomial error compensation. For this purpose, a polynomial correction algorithm is also implemented.

Results show that feedforward neural networks are able to reduce positional errors of electromagnetic tracking systems. Original errors are reduced in all acquired data sets and the network is able to predict model uncertainty. This is a great benefit since especially in healthcare applications, a wrong prediction may lead to serious consequences. Regarding transfer learning, neural networks outperform polynomials. The developed neural network reduces errors in any distorted scenario in case it is trained on laboratory data. Therefore, this network is a simple online calibration method. Additionally, neural networks make it easy to incorporate new information into the calculation like a metal distortion indicator which is recorded by the trakSTAR system. Except for transfer learning, polynomial regression often provides a positional error correction which is closer to the true positions than neural networks do. In addition, polynomials require less development effort: One of the main difficulties of neural networks is the time-consuming selection of suitable parameters. Both polynomials and neural networks require time-consuming training data acquisition which is not feasible in clinical daily life.

Gast Seminar

Zeit: 04.12.2019, 13:30-14:30

Ort: Raum 324 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referentin: Dr. Sandy Engelhardt

University of Applied Sciences in Mannheim

Titel: ”Deep Learning-based Image Analysis and Augmented Reality for Cardiology and Cardiac Surgery“

Abstract: Cardiac imaging improves on diagnosis of cardiovascular diseases by providing images at high spatiotemporal resolution. Manual evaluation of these time-series, however, is expensive and prone to biased and non-reproducible outcomes. At the same time, convolutional neural networks have shown outstanding performance in medical image processing. We present CNN methods that address named limitations by integrating segmentation, disease classification and motion modelling into a fully automatic processing pipeline. We will reflect on our winning contribution for MICCAI's 2017 Automated Cardiac Diagnosis Challenge (ACDC@STACOM workshop). Furthermore, we will show results from the translation of these methods to real heterogeneous clinical data and we provide information on the remaining performance gap.

In the second half of the talk, we will focus on computer-assisted surgery. One of our main research aims is to increase objectivism in surgery by the help of quantifications and better training concepts. We will present novel visualizations and measurement approaches of the mitral valve to improve performance in complex valve repair surgeries. Beyond that, we present enhanced tools for surgical training and unique concepts for phantoms based on 'Hyperrealism' (GANs). Towards the end of the talk, our initiative at research campus STIMULATE (University Magdeburg) is presented, where we mainly focus on echocardiography and tracking together with partners from industry.

Bio: Sandy Engelhardt is currently Post-Doc at University of Applied Sciences in Mannheim. She is PI of a DFG funded project and leader of the research group “Heart” at the research campus STIMULATE at Magdeburg University. The focus of STIMULATE are technologies for image guided minimally invasive methods in medicine. She studied Computational Visualistics at University of Koblenz-Landau and received the M.Sc. degree in 2012. From 2012 to 2016, she did her Ph.D. at the ”German Cancer ResearchCenter” (DKFZ) in Heidelberg in the division of ”Medical and Biological Informatics”. During her Ph.D., she was part of the Collaborative Research Center SFB/TRR 125 Cognition-guided surgery and contributed to an assistance system for reconstructive mitral valve surgery. The system was awarded with the MICCAI AE-CAI Best Paper Award in 2014. Her dissertation ”Computer-assisted Quantitative Mitral Valve Surgery” granted the BVM-Award 2017 for the best PhD thesis on the German Community. At MICCAI 2017, she together with a team won the ”Automated Cardiac Diagnosis Challenge”(ACDC) at the STACOM Workshop using a Deep Learning Approach. Her work mainly addresses topics of cardiology and computer-assisted cadiac surgery, thereby combining methods and technologies from the field of image segmentation, tracking systems and augmented reality.

Abschlussarbeiten

Zeit: 29.11.2019, 15:00

Ort: Raum 072 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Yannick Pflanzer (Betreuer: Martin Ritz)

Titel: ”Phenomenological Acquisition and Rendering of Optical Material Behavior for Entire 3D Objects“ (Bachelorarbeit)

Abstract: In the last few years, major improvements in 3D scanning and rendering technology have been accomplished. Especially the acquisition of surface appearance information has seen innovation thanks to phenomenological approaches for capturing lighting behavior. In this work, the current Bi-directional Texturing Function (BTF) and Approximate-BTF (ABTF) approaches were extended to allow for a greater depth of effects to be captured as well as the ability to reproduce entire 3D objects from different viewing angles.

The proposed Spherical Harmonic BTF (SHBTF) is able to model the captured surface appearance of objects by encoding all measured light samples into spherical harmonic coefficients, allowing for calculation of the surface appearance for any given light direction.

In contrast to the ABTF, an SHBTF can capture multiple views of the same object which enables it to efficiently reproduce anisotropic material properties and subsurface scattering in addition to the spatially varying effects captured by an ABTF.

The CultArc3D capturing setup used for all measurements is versatile enough to deliver view and light samples from a full hemisphere around an arbitrary object. It is now possible to capture entire 3D objects as opposed to many other BTF acquisition techniques. Challenges for the SH based lighting solution are ringing artifacts, growing stronger with rising SH bands.

Another challenge for a full 3D experience was the re-projection of camera images onto a 3D model, depending heavily on the camera hardware calibration. The SH based approach has the potential to produce compelling results given further optimizations of the SH and re-projection accuracy.

Zeit: 25.10.2019, 13:30

Ort: Raum 242 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Hendrik Pfeifer (Betreuer: Lena Cibulski)

Titel: ”A Visual Analytics Approach to Sensor Analysis for End-of-Line Testing“ (Masterarbeit)

Abstract: End-of-Line testing is the final step of modern production lines that assures the quality of produced units before they are shipped to customers. Automatically deciding between functional and defective units as well as classifying the type of defect are main objectives. In this thesis, a dataset consisting of three phase internal rotor engine simulations is used to outline opportunities and challenges of Visual Analytics for End-of-Line testing.

At first the simulation data is visually analyzed to understand the influence of the simulation input parameters. Afterwards features are extracted from the signals using discrete Fourier transform (DFT) and discreteWavelet transform (DWT) to represent the different simulations. Principal Component Analysis (PCA) is applied to further reduce the dimensionality of the data to finally apply K-Means to cluster the datasets and also perform a classification using a support vector machine (SVM). It is discussed which methods are beneficial for the End-of-Line testing domain and how they can be integrated to improve the overall testing process.

Zeit: 23.10.2019, 10:00

Ort: Raum 072 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Steffen Wirsching (Betreuer: Arjan Kuijper)

Titel: ”Object Localisation on Point Clouds with Deep Learning“ (Masterarbeit)

Abstract: A machine’s ability to recognize and pick objects from a bin is one of the most important applications in automation industry. This problem has been studied intensively and after all, bin-picking boils down to the task of localising objects on images / point clouds. While previous research has already achieved an improvement in the accuracy of bin-picking methods, only very few approaches focus on inference time as the central aspect to improve the task of bin-picking. Despite tremendous efforts by big companies to advance the field with challenges (e.g.the Amazon Picking Challenge), it is currently still lacking fast and super fast approaches.

This master thesis presents Fast Bin-Picking Convolutional Neural Network (FBCNN),a highly lightweight neural network architecture framework suitable for solving the task of detecting parts in a bin much faster than current solutions. In fact, a pose estimation of all parts in the bin can be generated in less than 6ms – which is more than sixty times faster than current state-of-the-art solutions. All that is needed for training is a CAD model of the parts and the bin so that synthetic training data can be generated and annotated automatically. This generation process is so fast that the network has a data set of infinite size available. Because of that, it benefits from aggressively discarding used data instead of splitting data into training and test sets.It is shown, that the features learned from synthetic data transfer to real data. FBCNN works on gray and depth images generated from camera(s) above the bin and directly outputs one 3D pose per detected part. It is built as a one-stage detector. To add to this, ablative studies on the details of neural networks are conducted in order to enable the adaptation of FBCNN to different applications. Finally, FBCNN is shown to work on both generated and real data and it features a hybrid training process to efficiently extract useful information from data samples.

GRIS Workshop

GRIS organisierte den Workshop “Visual Analytics in Healthcare” (VAHC)

Datum: 20. Oktober 2019

Ort: IEEE VIS, Vancouver, Canada

Paper Submission: https://new.precisionconference.com/submissions

Paper Deadline: 17. Juni 2019

URL: https://www.visualanalyticshealthcare.org/

Dr. Jürgen Bernard war einer der Organisatoren des VAHC-Workshops ”Visual Analytics in Medicine 2019“ am IEEE VIS, Vancouver, Kanada.

Visualisierungsexperten hatten die Möglichkeit, Experten aus verschiedenen Bereichen der Medizin zu treffen. Die Teilnehmer dieses Workshops diskutierten gemeinsam innovative visuell-interaktive Lösungen für die Analyse medizinischer und patientenbezogener Daten. Expertengruppen diskutierten die Bereiche und Anwendungen von Ärzten und tauschten sich über Best-Practice-Ansätze aus. In diesem Jahr wurde eine Zunahme von Lösungen erwartet, die explizit auf maschinelles Lernen und künstliche Intelligenztechniken zurückgreifen.

Teilnehmer an den Workshops konnten zwischen einem Full-Paper, einem Poster und einem Demo-Track wählen. Das Submission-System ist offline!

Awards at MICCAI 2019, 13-17 Oct 2019, Shenzhen, China

Es war ein sehr erfolgreiches Jahr für GRIS mit mehreren Auszeichnungen und sichtbarer Präsenz:

1. Auszeichnungen für die präoperative Planungsforschung durch Johannes Fauser

  • a. ”MICCAI Grad Student travel“ Auszeichnung
  • b. ”MICCAI workshop OR 2.0 best paper runner up“ Auszeichnung
  • c. ”MICCAI best presentation finalist“ (15 von 783)

2. Community Service

  • a. ”Outstanding reviewer“ Auszeichnung an David Kügler und Johannes Fauser
  • b. Anirban Mukhopadhyay leitete durch die ”CAI oral session“ (auch als Area Chair)

3. Educational award für Camila Gonzalez

  • a. 3. Preis in ”MICCAI Educational Challenge“

4. GAN Tutorial mitorganisiert mit Arjan Kuijper

  • a. 200+ Teilnehmer

Abschlussarbeiten

Zeit: 24.09.2019, 11:00

Ort: Raum 324 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Angeelina Rajkarnikar (Betreuer: Anirban Mukhopadhyay)

Titel: ”Practical Defense against Medical Adversarial Examples“ (Masterarbeit)

Abstract:

This thesis aims to provide practical solutions for two different problems using Machine Learning and Deep Learning (DL) techniques in Healthcare applications: a) A practical defense against adversarial attacks so that DL can be used without vulnerabilities. b) A practical framework for evaluating Machine Learning (ML) algorithms on spectroscopic data. With increasing benefits of using deep learning in clinical applications, the vulnerabilities of Deep Neural Networks to adversarial examples are also surfacing. Adversarial examples are images that are injected with imperceptible perturbations that results in it’s misclassification in image classification problems. Healthcare, being a safety critical field is vulnerable to adversarial attacks, yet there are limited research on practical defenses on making the network robust against these attacks. We propose a defense strategy against adversarial attacks on deep learning classification models used in the field of medical diagnosis by adding random noise at inference time. Empirically, we evaluate the effectiveness of this simple defense strategy against single-step and iterative adversarial attacks. We demonstrate our technique through experiments, by introducing randomness, the accuracy of attacked images are increased. This techniques shows an significant increase in accuracy from 2.2% to 98.5% for MNIST, 8% to 88.08% for CIFAR-10 and 2.4% to 76.81% for HAM10000. This approach makes no assumptions about the model architecture. Hence it is simple to implement, yet effective in developing more robust deep learning models.

Infrared (IR) spectroscopy is a non-invasive analytical method successfully used for multicomponent analysis of bio-fluid samples. We develop a ML framework with various multivariate calibration algorithms to predict the concentrations of components in the dialysate using IR spectroscopy. We analytically show that the obtained results of correlation between predicted and clinical reference methodology readings for 5 components are reasonable. Keywords: Adversarial attack, adversarial defense, FTIR Spectroscopy.

Zeit: 17.09.2019, 11:00

Ort: Raum 220 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Daniel Alte (Betreuer: Anirban Mukhopadhyay)

Titel: ”Spatial interpretation of DNN-Classifiers in medical imaging“ (Masterarbeit)

Abstract:

A shortage of radiologists in both industrialized countries and developing countries leads to delayed or absent treatments of diseases. Modern medical artificial intelligence systems can potentially assist radiologists to achieve more efficient detection and diagnosis of diseases.

However, such systems rely on deep neural networks, which are mostly black boxes. Especially the clinical sector requires interpretable systems. Recent work introduces tractable uncertainty estimates using dropout (Monte Carlo Dropout). For classification tasks, this could indicate a radiologist, how confident a network is in its prediction. Uncertainty estimates answer the question of if a network ”knows“, whether it is predicting falsely. A different way of interpreting a Convolutional Neural Network are Class Activation Maps, a visualization technique showing discriminative spatial features for a given classification. This thesis investigates, whether combining Class Activation Maps and Monte Carlo Dropout leads to obtaining spatial uncertainty estimates. A correlation analysis shows that the predictive entropy can separate spatial uncertainties in two clusters. The proposed method is evaluated qualitatively. It is observed that redundant features can cause high spatial uncertainties. Furthermore, we show qualitatively that the proposed method finds spatial uncertainties. Additionally, it has to be noted that our approach treats each pixel of the uncertainty map independently from all other pixels. Future work can investigate the maps with dependent pixels.

Zeit: 11.09.2019, 14:00

Ort: Raum 324 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Martin Müller (Betreuer: Marcel Wunderlich)

Titel: ”Modellbasiertes visuell-interaktives System zur Hypothesenbildung bzgl. Übertragungswegen von Infektionen im Krankenhaus“ (Masterarbeit)

Abstract: Jährlich werden 400.000 – 600.000 Patienten in Deutschland während eines Krankhausaufenthaltes (“nosokomial“) infiziert [SHK13, GG08]. Es versterben ca. 10.000 – 15.000 Patienten in Deutschland jährlich an den Folgen dieser Infektionen [GG08]. Im Rahmen dieser Masterarbeit wird ein visuell-interaktives System entwickelt, welches Hypothesen zu möglichen Übertragungswegen zwischen Patienten generiert. Zudem können Ärzte und Mikrobiologen, die Nutzer des Systems, dieses auch nutzen, um mögliche bisher nicht identifizierte Infektionen von Patienten aufzudecken und Infektionshäufungen räumlicher und zeitlicher Natur zu finden. Dadurch unterstützt das System die Anwender dabei, die Ausbreitung von Krankenhausinfektionen einzudämmen. Die Anforderungen, die das System zu erfüllen hat, wurden in Kooperation mit Ärzten und Mikrobiologen der Uniklinik Heidelberg entwickelt. Die Datenbasis, hauptsächlich bestehend aus Aufenthaltsdaten und Labortestresultaten, stammt ebenfalls von der Uniklinik Heidelberg. Das System besteht aus zwei Teilen: einem Modell, welches den Gesundheitszustand von Patienten und Übertragungswege zwischen diesen modelliert, sowie einer interaktiven Visualisierung, die Modellresultate für Ärzte und Mikrobiologen angemessen visualisiert. Um den praktischen Mehrwert des kreierten Systems zu demonstrieren wird ein Use-Case mit einem Mikrobiologen der Uniklinik Heidelberg entwickelt. Es zeigt sich, dass dieses System Ärzte wirksam bei der Identifizierung von Infektionsschwerpunkten, bei der Analyse von Übertragungswege und bei der Identifizierung von infizierten Patienten unterstützen kann.

Gastvortrag

Zeit: 09.09.2019, 14:30 Uhr

Ort: Raum 073 im Fraunhofer IGD Fraunhoferstrasse 5, S3|05

Referent: Prof. Dr. Tobias Schreck,

Institute of Computer Graphics and Knowledge Visualization at TU Graz, Austria

Titel: ”Visual Analytics Approaches for Data Exploration: Visual Cluster Analysis, Visual Pattern Search, and Immersive Analytics“

Abstract: Visual Analytics approaches support users in interactive data exploration and pattern discovery, relying on data visualization integrated with steerable data analysis algorithms. After a brief introduction to basic ideas of Visual Analytics, we discuss examples of Visual Analytics research from our current work. First, we discuss interactive visual clustering for exploration of time series and trajectory data. Then, we discuss approaches for retrieval, comparison and modeling of visual patterns in high-dimensional data. Thirdly, we discuss ongoing work in immersive analytics of movement data captured in VR-based training applications. We close by highlighting research opportunities, including user guidance and eye tracking as an analytic interaction modality.

CV: Tobias Schreck is a Professor with the Institute of Computer Graphics and Knowledge Visualization at TU Graz, Austria. Between 2011 and 2015, he was an Assistant Professor with the Data Analysis and Visualization Group at University of Konstanz, Germany. Between 2007 and 2011, he was a Postdoc researcher and research group leader with the Interactive-Graphics Systems Group at TU Darmstadt, Germany. Tobias Schreck obtained a PhD in Computer Science in 2006 from the University of Konstanz. His research interests are in Visual Analytics and applied 3D Object Retrieval. Tobias Schreck has served as a papers co-chair for the IEEE VIS Conference on Visual Analytics Science and Technology (VAST) in 2018 and 2017. For more information, please see http://www.cgv.tugraz.at/schreck/.

Gastvortrag

Zeit: 05.08.2019, 10:30-11:30 Uhr

Ort: Raum S101|A2, Universitätszentrum, karo 5, Karolinenplatz 5

Referent: Dr. Ilkay Oksuz

King's College London Biomedical Engineering Department

Titel: ”Automatic Quality Assessment of Cardiac MRI using Deep Learning Techniques“

Abstract: Cardiovascular disease (CAD) is the major cause of mortality in the world. Recently, Cardiovascular Magnetic Resonance (CMR) techniques have gained ground in diagnosis of cardiovascular disease and good quality of such MR images is a prerequisite for the success of subsequent image analysis pipelines. Quality assessment of medical images is therefore an essential activity and for large population studies such as the UK Biobank (UKBB), manual identification of artefacts such as those caused by unanticipated motion is tedious and time-consuming. In this talk, recent work on detection of wrong cardiac planning and cardiac motion artefacts using deep learning techniques will be described. The details of deep learning architectures and machine learning methodologies will be given with a certain focus on synthetic k-space corruption and curriculum learning techniques. In the last part of the talk, the mechanisms to correct image artefacts will be discussed alongside with their influence on achieving high segmentation accuracy.

Bio: Dr. Ilkay Oksuz is currently a Research Associate in King's College London Biomedical Engineering Department. His current research interests are in medical image segmentation, medical image registration and machine learning, with a focus on the automated analysis and quality control of cardiac MR. He studied for a PhD at the IMT Institute for Advanced Studies Lucca on Computer, Decision, and Systems Science under the supervision of Prof Sotirios Tsaftaris. His PhD thesis focused on joint registration and segmentation of the myocardium region in MR sequences. He joined the Diagnostic Radiology Group at Yale University in 2015 for 10 months as a Postgraduate Fellow, where he worked under the mentorship of Prof Xenios Papademetris. He also worked at the University of Edinburgh Institute for Digital Communications department for six months in 2017.

Referenzen:

https://link.springer.com/chapter/10.1007/978-3-030-00129-2_3

https://link.springer.com/chapter/10.1007/978-3-030-00928-1_29

https://www.sciencedirect.com/science/article/pii/S1361841518306765

https://ieeexplore.ieee.org/abstract/document/8363616

https://openreview.net/forum?id=BkgjbQ30yN

Poster Präsentation

Deep Generative Models 2019 (by MEC-Lab@GRIS)

Zeit: 16.07.2019, 10:00-11:00 Uhr

Ort: Raum 073 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Studierende der LV Deep Generative Models

Themenbeispiele:

1. Food Interpolator – fluid transition between pizza and burger

2. Generating Instagram images from hashtags

3. Interpolate over the space from the street view house numbers (SVHN) dataset

Abschlussarbeiten

Zeit: 27.05.2019, 16:00

Ort: Raum 324 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Markus Lehmann (Betreuer: Jürgen Bernard)

Titel: ”Visual-Interactive Combination of Selection Strategies to Improve Data Labeling Processes“ (Masterarbeit)

Abstract:Labeling training data is an important task in Machine Learning for building effective and efficient classifiers. There are different approaches to gather labeled instances for a particular data set. The two most important fields of strategies are Active Learning and Visual-Interactive Labeling. In previous work, these strategies were examined and compared, resulting in a set of atomic labeling strategies. Additionally, a quasi-optimal strategy was analyzed in order to infer knowledge from its behavior. This investigation resulted in two main insights. First, the labeling process consists of different phases. Second, the performance of a strategy depends on the data set and its characteristics.

In this work, we propose a toolkit which enables users to create novel labeling strategies. First, we present multiple visual interfaces users can employ to examine the space of existing algorithms. Then, we introduce a definition of ensembles users can build upon in order to combine existing strategies to novel strategies. Multiple methods to measure the quality of labeling strategies are provided to users, enabling them to examine the gap between their strategies and existing strategies. The different phases of the labeling process are included in the toolkit in order to allow users to always apply the most appropriate strategy in each phase. During the entire process, users are supported by automated guidance in the improvement of their strategies.

We evaluate our concept from different perspectives in order to assess its quality. Overall, we observe that our approach enables users to build ensemble strategies which outperform existing strategies. The insights from this work can be applied to develop novel concepts towards ensemble building as well as to improve the generalization of strategies to other data sets.

Zeit: 22.05.2019, 10:00

Ort: Raum 072 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Moritz Sachs (Betreuer: Matthias Unbescheiden)

Titel: ”Automatisierte Geschäftsmodellanalysen mit Deep Neural Networks“ (Masterarbeit)

Abstract:Ein Hauptkriterium für das Investment eines Venture Capital Fonds in ein Start-up ist dessen Geschäftsmodell. Dieses ist im Businessplan enthalten. Das Screening, sowie die Analyse der eingereichten Businesspläne, erfolgt bei den meisten Venture Capital Fonds überwiegend durch Menschen.

Mit der vorliegenden Arbeit wird untersucht, inwieweit die Analyse der in den Businessplänen enthaltenen Geschäftsmodelle mit Hilfe von Deep Neural Networks automatisiert werden kann. Ziel war die Entwicklung eines Prototypen, der die in den Businessplänen enthaltenen Geschäftsmodelle automatisch extrahiert und in das Metamodell Startup Navigator überführt.

Dem Knowledge Discovery in Databases Prozess folgend wurden hierfür die Businesspläne eines Venture Capital Fonds aufbereitet und damit ein tiefes Convolutional Neural Network, der Multilabel k-Nearest Neighbour Algorithmus, sowie eine Support Vector Machine mit Naive Bayes Features trainiert.

Die Ergebnisse des entwickelten Prototypen zeigen, dass die in den Businessplänen enthaltenen Geschäftsmodelle automatisch extrahiert und in das Metamodell Startup Navigator überführt werden können. Es erscheint plausibel, dass mit mehr Trainingsdaten und einer intensiveren Hyperparameteroptimierung die Korrektklassifizierungsrate verbessert werden kann, sodass der Prototyp zum Aufbau eines Geschäftsmodellkorpus genutzt werden könnte.

Zeit: 30.04.2019, 10:00

Ort: Raum 103 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Moritz Matthiesen (Betreuer: Pavel Rojtberg)

Titel: ”Interpolation von Kalibrierdaten für Zoom und Autofokus Kameras“ (Bachelorarbeit)

Abstract: In dieser Arbeit wird das Problem betrachtet, dass für jede neue Kameraeinstellung eine neue Kalibrierung vorgenommen werden muss.

Ziel dabei ist Kalibrierdaten an bestimmten Kameraeinstellungen zu erstellen, um mithilfe von diesen die Kalibrierdaten von anderen Kameraeinstellungen herzuleiten. Dabei werden die Kalibrierdaten betrachtet und es wird versucht Beziehungen zwischen den einzelnen Parametern der Kalibrierung herzuleiten. Um diese zu ermitteln wird zwischen verschiedenen Parametern der Kalibrierung interpoliert.

Zeit: 29.04.2019, 15:00 Uhr

Ort: Raum 324 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Ali Jabhe (Betreuer: David Kügler)

Titel: ”Physical World Attacks in Medical Imaging“ (Bacheloarbeit)

Abstract: The methodology and the acquisition of images for the attacks on dermoscopy tackles the question of whether Deep-Learning Systems can be attacked by a malicious attacker without changing anything on the Deep-Learning side. That mean only changes on the physical world are allowed. This problem is an extension of the ”adversarial attack“ concept, but with some twist.

Zeit: 25.04.2019, 14:00 Uhr

Ort: Raum 324 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Heiko Reinemuth (Betreuer: Jürgen Bernard)

Titel: ”Visual-Interactive Labeling of Multivariate Time Series to Support Semi-Supervised Machine Learning“ (Masterarbeit)

Abstract: Labeling of multivariate time series is an essential requirement of the data-centric decision-making processes in many time-oriented application domains. The basic idea of labeling is to assign (semantic) meaning to specific sections or time steps of the time series and to the time series as a whole, accordingly. Hence, weather phenomena can be characterized, EEG signals can be studied, or movement patterns can be marked in sensor data.

In the context of this work a visual-interactive labeling tool was developed that allows non-expert users to assign semantic meaning to any multivariate time series in an effective and efficient way. Enabling experts as well as non-experts to label multivariate time series in a visual-interactive way has never been proposed in the information visualization and visual analytics research communities before. This thesis combines active learning methods, a visual analytics approach, and novel visual-interactive interfaces to achieve an intuitive data exploration and labeling process for users. Visual guidance based on data analysis and model-based predictions empowers users to select and label meaningful instances from the time series. As a result, the human-centered labeling process is enhanced by algorithmic support, leading to a semi-supervised labeling endeavor combining strengths of both humans and machines.

Zeit: 17.04.2019, 10:00 Uhr

Ort: Raum 242 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Johann Reinhard (Betreuer: Alan Brunton)

Titel: ”Efficient streaming sample-based surface triangulation of voxel data“ (Masterarbeit)

Abstract: Voxel-based discrete representations of 3-dimensional data are widely used in several fields of graphical computing, for instance in the 3D printing driver Cuttlefish. For commonly used techniques, such as the marching cubes algorithm, the creation of a polygonal/polyhedral mesh representation of the used voxel data at high resolutions can become time-consuming and result in meshes with excessive numbers of vertices, which nonetheless introduce ”staircase“ artifacts relative to the desired geometry. It is then often necessary to use additional post-processing steps, such as mesh decimation, at the expense of additional computational effort and possible inaccuracies regarding the representation of the original shape.

The goal of this thesis is to simultaneously address all three of these issues, proposing an efficient technique to generate low-polygon meshes, which accurately represent the object’s shape. The intended technique is based on sampling the surface at regions of high curvature using, for example, an importance sampling technique, although different techniques will be explored. A comparison will be made between per-slice and per-chunk sampling (i.e. consider only a single slice or a whole chunk of slices when deciding where to place samples). The samples are to be mapped to a parametric, planar space, allowing to efficiently triangulate the sampled points. The necessity of additional post-processing steps in the parametric or reprojected object space will be assessed. The developed techniques will be implemented, integrated into Cuttlefish and evaluated based on comparisons to standard techniques such as marching cubes or Marching Tetrahedra using the above three measures: efficiency (time and memory), number of polygons in the output, and accuracy. Defining a measure of the accuracy of the output and computing it is a further aspect of the thesis, where at least the Hausdorff distance and collinearity of the surface normals will be measured in order to quantify the mesh quality.

Zeit: 20.03.2019, 15:00 Uhr

Ort: Raum 073 im Fraunhofer IGD, Fraunhoferstrasse 5, S3|05

Referent: Alexander Distergoft (Betreuer: Anirban Mukhopadhyay)

Titlel: ”Interpreting Adversarial Examples in Medical Imaging" (Masterarbeit)

Abstract: Deep neural networks (DNNs) have been achieving high accuracy on many important tasks like image classification, detection or segmentation. Yet, recent discoveries have shown a high degree of susceptibility for these deep-learning algorithms under attack. DNNs seem to be vulnerable to small amounts of non-random noise, created by perturbing the input to output mapping of the network. These perturbations can severely affect the performance of DNNs and thus endanger systems where such models are employed.

The purpose of this thesis is to examine adversarial examples in clinical settings, be it digitally created or physical ones. For this reason we studied the performance of DNNs under the following three attack scenarios:

1. We hypothesize that adversarial examples might occur from incorrect mapping of the image space to the lower dimensional generation manifold. The hypothesis is tested by creating a proxy task of a pose estimation of surgical tools in its simplest form. For this we define a clear decision boundary. We use exhaustive search on a synthetic toy dataset to localize possible reasons of successful one-pixel-attacks in image space.

2. We design a small scale prospective evaluation on how Deep-learning (DL) dermoscopy systems perform under physical world attacks. The publicly available Physical Attacks on Dermoscopy Dataset (PADv1) is used for this evaluation. The introduced susceptibility and robustness values reveal that such attacks lead to accuracy loss across popular state-of-the-art DL-architectures.

3. As a pilot study to understand the vulnerabilities of DNNs that perform under regression tasks we design a set of auxiliary tasks that are used to create adversarial examples for non-classification-models. We train auxiliary networks on augmented datasets to satisfy the defined auxiliary tasks and create adversarial examples that might influence the decision of a regression model without knowing about the underlying system or hyperparameters.