Events
Talk Reconstructing 3D Human Avatars from Monocular Images
28.08.2019 13:00
Informationzentrum, Seminarraum G30
Speaker(s): Thiemo Alldieck
Modeling 3D virtual humans has been an active field of research over the last decades. It plays a fundamental role for many applications, such as movie production, sports and medical sciences, or human-computer interaction. Early works focus on artist-driven modeling or utilize expensive scanning equipment. In contrast, our goal is the fully automatic acquisition of personalized avatars using low-cost monocular video cameras only. In this dissertation, we show fundamental advances in 3D human reconstruction from monocular images. We solve this challenging task by developing methods that effectively fuse information from multiple points in time and realistically complete reconstructions from sparse observations. Given a video or only a single photograph of a person in motion, we reconstruct, for the first time, not only his or her 3D pose but the full 3D shape including the face, hair, and clothing. We investigate various aspects of monocular image and video-based 3D human reconstruction and demonstrate both straight-forward and sophisticated reconstruction methods focused on accuracy, simplicity, usability, and visual fidelity. During extensive evaluations, we give insights into important parameters, reconstruction quality, and the robustness of the methods. For the first time, our methods enable camera-based, easy-to-use self-digitization for exciting new applications like, for example, telepresence or virtual try-on for online fashion shopping.
Talk How much “nature” is in the image? The role of lower-level processed image properties on the processing and evaluation of faces, artworks, and environments
19.08.2019 15:00
Informationzentrum, Seminarraum G30
Speaker(s): Claudia Menzel
The human visual system is adapted to natural scenes and their characteristics, which leads to an efficient / fluent processing of such scenes. Interestingly, art and other aesthetically pleasing images share such natural image properties. Thus, these properties inherent in natural scenes are associated with beauty. In my talk, I will present computational, behavioural, and neurophysiological data on the relationship of natural image properties and the processing and evaluation of three stimulus categories: faces, artworks, and environmental scenes. In the first part of my talk, I will present a series of studies showing that natural image properties beneficially influence face processing and perceived facial attractiveness. In the second part, I will present studies on the role of image properties for the fast and automatic detection of artistic composition in artworks. In the third part, I will come back to natural scenes and present current data on the role of image properties for the evaluation of and health effects evoked by nature and urban environments. Overall, the presented work will demonstrate how image properties inherent to natural scenes influence the processing and evaluation of various image categories.
Talk MA-Talk: Augmented Reality in Optics Laboratories
29.07.2019 13:00
Informatikzentrum, Seminarraum G30
Speaker(s): Nils Wendorff
Talk MA-Talk: Quality Metric for Scatter Plots Based on Human Perception using CNNs
05.07.2019 13:00
Informatikzentrum, Seminarraum G30
Speaker(s): Yuxin Zou
Symposium Real VR - Importing the Real World into Immersive VR and Optimizing the Perceptual Experience of Head-Mounted Displays
30.06.2019 18:00
- 03.07.2019 13:00
Dagstuhl
Chair(s): Marcus Magnor
Motivated by the advent of mass-market VR headsets, our Dagstuhl Seminar addresses the scientific and engineering challenges that need to be overcome in order to experience omni-directional video recordings of the real world with the sense of stereoscopic, full-parallax immersion as can be provided by today’s head-mounted displays.
Talk Reconstructing 3D Human Avatars from Monocular Images
27.06.2019 12:00
Google Zürich
Speaker(s): Thiemo Alldieck
Statistical 3D human body models have helped us to better understand human shape and motion and already enabled exciting new applications. However, if we want to learn detailed, personalized, and clothed models of human shape, motion, and dynamics, we require new approaches that learn from ubiquitous data such as plain RGB-images and video. I will discuss recent advances in personalized body shape and clothing estimation from monocular video, from a few frames, and even from a single image. We developed effective methods to learn detailed avatars without the need for expensive scanning equipment. These methods are easy to use and enable personalized avatar creation for example for VR and AR applications. I will conclude my talk by outlining the next challenges in human shape reconstruction.
Talk Scene-Space Video Processing
17.05.2019 13:00
Informationzentrum, Seminarraum G30
Speaker(s): Felix Klose
quuxLogic Solutions GmbH
Nearly all image based video processing methods require correspondences to be established in between input images. Finding image correspondences is an area of active research for many years, and is known to be a notoriously difficult problem, due to its ill-posed nature and large number of unknowns. If the input data is not only a few images, but single or multiple video streams, the amount of data to be processed poses an additional challenge. This thesis provides a insight into commonly applied constraints to correspondence estimation in 2D, 3D and 4D and the applications that become possible from using the different data modalities. It is shown how free-viewpoint stereoscopic 3D renderings of multi-view recordings can be generated by purely using 2D correspondences. To accomplish this in the required high output quality, a simple interaction paradigm is introduced, that allows users to correct correspondences where the automatic results are not yet of sufficient quality. An approach to compute full quasi-dense scene-flow for short time intervals is presented. This work then shows how new applications become possible, when the inaccuracies in scene estimation are leveraged and the full redundancy of the available visual data is exploited. And highlights the performance implications of using the large amounts of available data.
Talk Tell Me How You Look and I'll Tell You How You Move
26.04.2019 09:15
Tampere University, Tampere, Finland
Speaker(s): Thiemo Alldieck
Talk at CIVIT Tech Day: Motion Deconstructed
[Video]
Human body shape and motion estimation are two sides of the same coin. To be able to fully understand human motion from monocular imagery, we need to understand the shapes of the tracked subjects, too. In my talk, I will motivate why we need better shapes for better tracking. I will demonstrate how 3D bodies helped us to understand human motion better and where these models find their limits. If we want to learn rich models of human shape, motion, and dynamics, we require new approaches that learn from ubiquitous data such as plain RGB-images and video. I continue with discussing recent advances in personalized body shape estimation from monocular video, from a few frames, and even from a single image. We developed effective methods to learn detailed avatars without the need for expensive scanning equipment. These methods are easy to use and enable various VR and AR applications. I will conclude my talk by outlining the next challenges in human shape reconstruction and how this potentially affects human motion estimation.
Conference Computational Visual Media Conference (CVM)
24.04.2019 00:00
- 26.04.2019 00:00
University of Bath, UK
Chair(s): Marcus Magnor
Talk BA-Talk: Exploring high-dimensional data with Parallel Coordinates
15.04.2019 13:00
Informationzentrum, Seminarraum G30
Speaker(s): Waldemar Gorbunov
Talk Data-driven estimation and neuro-physiological assessment of perceived visual quality
12.04.2019 13:00
Informatikzentrum IZ, Seminarraum G30
Speaker(s): Sebastian Bosse
FhG-HHI Berlin
Perceptual quality is a key aspect of modern multimedia systems. Nevertheless, the questions of how to computationally estimate perceived quality, how to incorporate computational models of quality in technical systems and how to reliably assess the true quality as perceived by humans remain answered unsatisfactorily. In my talk I will outline novel neural network-based and end-to-end optimized models for visual quality estimation and show how such models can be used in practice, i.e. for perceptually optimized bit allocation in state-of-the-art video compression schemes. The reliability of these data-driven models rely on the availability of large quality annotated data sets. For generating such data sets, conventional psychophysical approaches inherently suffer from several crucial drawback. My talk will discuss how neuro-physiological quality assessment may overcome these flaws and pave the way towards reliable assessment of perceptual quality.
Talk Teamprojekt-Abschluss: A Look Through Floyd Lawton's Mask
01.04.2019 13:00
Dome (Aufnahmestudio & Visualisierungslabor)
Präsentation der Ergebnisse des studentischen Teamprojekts.
Talk Praktikum-Abschluss: World Builder VR Toolkit Extended
29.03.2019 13:00
Dome (Aufnahmestudio & Visualisierungslabor)
Präsentation der Ergebnisse des studentischen Computergraphik Praktikums.
(Ein Folgeprojekt vom Teamprojekt WS '17/18)
Talk MA-Talk: Human Pose Estimation in Augmented Reality Scenarios
22.03.2019 13:00
Informatikzentrum, Seminarraum G30
Speaker(s): Alexander Meyer
Talk MA-Talk: Evaluation von Visualisierungstechniken unter Nutzung von Eye-Tracking und EEG
01.03.2019 13:00
Informatikzentrum, Seminarraum G30
Speaker(s): Anton Antonenko
Talk BA-Talk: Analyse- und Berechnungsverfahren zur Auswertung eines pixelierten Detektors mit einem Gleitschatten-Filter
15.02.2019 13:00
Seminarraum G30
Speaker(s): Felix Lehner
Talk Turning Reality into Virtual Reality
30.11.2018 11:00
FhG Heinrich Hertz Institut Berlin
Speaker(s): Marcus Magnor
Current endeavors towards immersive visual entertainment are still almost entirely based on 3D graphics-generated content, limiting application scenarios to virtual worlds only. The reason is that in order to provide for stereo vision and ego-motion parallax, two essential ingredients for genuine visual immersion perception, the scene must be rendered in real-time from arbitrary vantage points. While this can be easily accomplished with 3D graphics via standard GPU rendering, it is not at all straight-forward to do the same from conventional video footage acquired of real-world events. In my talk I will outline avenues of research toward enabling the immersive experience of real-world recordings and how to enhance the immersive viewing experience by taking perceptual issues into account.
Talk MA-Talk: Learning Optical Flow from Long-Exposure Images
23.11.2018 11:00
Seminarraum G30
Speaker(s): Moritz Kappel
Recently, Convolutional Neural Networks (CNNs) have been applied to a wide range of tasks including classification, detection, segmentation
and optical flow. Despite convincing results the overall capability of CNNs depends on the utilized training data for supervised learning. This
is especially restricting for optical flow estimation as obtaining ground truth optical flow is infeasible for the bulk number of real world scenes
required for training. As a consequence approaches for unsupervised learning gained more interest.
In this thesis, a CNN was designed to learn the Optical Flow between two images, consisting of the backward and forward motion for
every pixel. Furthermore, an occlusion moment was learned, at which a pixel becomes visible in the target image. For simplicity only
linear motion was considered, as this is sufficient to approximate more complex motion with small time-steps. Unlike traditional supervised Optical Flow networks, the CNN in this thesis was not trained with ground truth optical flow but with long-exposure images between the two images.
Talk BA-Talk: GPU-basierte Bildverarbeitung zur Analyse neuronaler Aktivität in Zebra fischlarven
05.11.2018 13:00
Seminarraum G30
Speaker(s): Immanuel Becker
Bei der Analyse neuronaler Aktivität von lebenden Zebra schlarven mittels Lichtblattmikroskopie werden Auflösungen erreicht, durch die einzelne Neuronen sichtbar werden. Für jede Aufnahme durch das Mikroskop wird der Fisch schichtweise durchleuchtet, sodass ein 3D Bild seines Kopfes entsteht.
Viele solcher 3D Aufnahmen über die Zeit bilden eine große Datenmenge deren Verarbeitung viel Zeit beansprucht. Das Ziel dieser Arbeit ist, GPU-basierte Beschleunigungen zu präsentieren, die die Ausführungszeit von kritischen Stellen der Verarbeitungsroutine verbessern.
Talk Digital Personality and The Emotional Onion
19.10.2018 13:00
IZ G30
Speaker(s): Susana Castillo
BTU Cottbus
Communication is inherent to life. During our daily life we convey information verbally and nonverbally. We can learn a lot about someone by watching their facial expressions and body language. Used to this kind of interaction with their interlocutors, people tend to personify machines, thus, harnessing these
aspects of non-verbal communication can lend virtual agents greater depth and realism, by giving them the ability to actually produce social information. But, in order to achieve these goals, we require a sound understanding of the relationship between cognition and expressive behavior.
This talk explores the multi-layered nature of communication and introduces an extended traditional word-based methodology to use both actual videos and motion capture data to extract the semantic/cognitive space of facial expressions. These recovered spaces can capture the full range of facial communication and are very suitable for semantic-driven facial animation.
Nevertheless, to provide virtual agents with full human-like communicative capabilities, natural facial expressions are not enough. We would like to impregnate all emotions with coherent deviations from the ’neutral’ emotions that can be perceived as series of personality traits which will result in the classification
of the agent as a particular kind of individual (e.g. an aggressive character or a cheerful one). Thus, this talk also will explore how to provide the agent with a personality that can be appreciated by the interlocutor.
In summary, this talk introduces how to find a characterization of the dynamic structure of facial expressions in two different levels. The high level focuses on obtaining the higher-level semantic structure underlying the facial expression space. And, the low-level contains the cues that allow us to replicate any expression by defining the particular movements that conform it, so to say, allow us to map the semantic space and the facial expressions. Last but not least, we will comment the implicit indication of the mapping between this semantic space for facial expressions and the personality space.
Talk MA-Talk: Material Parameter Extraction From Photos
12.10.2018 11:00
Seminarraum G30
Speaker(s): Sascha Fricke
The extraction of material parameters from photographs involves a process generally known as inverse rendering. This process is highly ambiguous and thus difficult to solve. Physically based shading and rendering systems allow the synthesis of believable images from such material parameters. Inverting this process directly is unfortunately not possible. Reformulating this into an optimization problem would allow the estimation of these properties by Gradient Descent. Derivatives of the rendering process with respect to the material parameters are however difficult to find due to discontinuous integrals and even if approximate derivatives are found, the ambiguity of the problem still leads to local minima that do not correspond to the desired material properties.
In the last years, convolutional neural network based systems have been shown to be capable of extracting these material parameters under certain restrictions. Unfortunately, the resulting extracted parameters are either in a parametrization, which is not useful for further use in game or film productions, or are not spatially varying and hence don’t contain enough information. Systems, that extract spatially varying parameters in the desired parametrization do exist, but are mostly only able to extract repeating, regular texture patterns from flat surfaces under simple lighting conditions.
In this thesis, a physically based, differentiable rendering system approach is presented and evaluated that is coupled with an initial, coarse estimation from a convolutional neural network. This system is able to improve the quality of estimations of both individual approaches and can be used in a general and messy real world setting to extract reflectance properties from photos that can be used as textures on triangle based geometry in real time or offline rendering systems.
Talk BA-Talk: Improving Training of Convolutional Neural Networks using Visualization Techniques
17.08.2018 13:30
Seminarraum G30
Speaker(s): Jann-Ole Henningson
Talk MA-Talk: Analyzing the Performance of Deep Neural Networks on Synthetic Training Data
17.08.2018 13:00
Seminarraum G30
Speaker(s): Meng Wang
Talk MA-Talk: An Augmented Reality Framework to Support the Implementation of Educational Applications
22.06.2018 13:00
- 22.06.2018 13:30
G30
Speaker(s): Manuel Behlen