|
max planck institut
informatik 
|
|
max planck institut
informatik
|
As you could see from my projects page, there is a variety of very exciting projects on the boundary between computer vision and computer graphics that I am interested in. I am always looking for students who are in pushing the limits of technology further within a Master or Bachelor thesis. I always have a variety of Master thesis topics available from the following areas of research. If you you want to get more detailed information, please get in touch with me via Email.
We have built some of the first algorithms in the literature to capture highly detailed dynamic geometry, motion and appearance models of human actors in general apparel from only a handful of multi-view video recordings, see link for an example. In contrast to commercial marker-based motion capture systems, our algorithms don't require the performers to wear any optical markers. Our methods greatly exceed the abilities of any commercially available system. There is also a very high interest from movie, game and motion capture companies in this kind of research. If you are interested in doing your Master thesis in this area, you will be able to use all the methods that we have available, and also get access to our state-of-the-art multi-view recording studio which features a variety of video camera systems, as well as a full-body laser scanner. There is a multitude of remaining open challenges in the field of performance capture and I would be happy to tell you how you could help us solving them.
 |  |
 |  |  |
A Time-of-flight camera is a new active type of sensor that allows real-time depth recording. The advantage over other passive reconstruction methods, like stereo, is that they typically perform very well even if there is sparse or no texture in the scene. We have several of these very new cameras available and also run one of the first joint multi-view depth and video camera systems. Time-of-flight cameras have great potential to enable a variety of new applications, in particular in the fields of real-time interaction, virtual and augmented reality, robotics, environment perception and understanding, real-time motion capture, computational photography and videography, 3D video etc. There are several algorithmic challenges involved, in particular around the question how to understand and model the non-trivial noise characteristics, and how to fuse information from video and depth cameras for better reconstruction. In our prvious research, we already looked at some of these problems, but there is a lot more room for innovation. I would be happy to discuss available research projects with you, in particular in the field of real-time 3D interaction and motion capture. The following are a few images from our research on 3D superresolution, ad sensor fusion. Please also check my projects page.
 |  |  |
This is a fairly new project which we recently started. The guiding idea is to analyze and process community video collections in order to automatically extract higher level information (about places, objects etc.) that are seen in the videos. The overall goal is to capitalize on the ever growing amount of casually capture video and develop tools which allow the everyday user to benefit from the huge amount of video on the web when creating his own video footage. We developed a variety of ideas for cool projects in this context. Each of them require solutions to hard problems from a variety of problem fields, including computer vision, semantic analysis etc. There is, however, also always a creative component that goes along with researching the core algorithmic problems. I would be happy to discuss available topics in this area with you during a personal meeting.
In this fairly new research project we strive to create more believable avatars for networked virtual environments. An avatar is a virtual human that can be controlled by a user in order to enable interaction in networked virtual environments or games. Currently, the appearance of avatars is often rather "wooden", and the means of interaction with the avatar in the virtual world are rather limited.
One of our goals is to equip avatars with the ability to perform believable gestures. Virtual environments exist in which the avatar "speak" can speak what the user speaks into the PC's microphone. Unfortunately, the avatar cannot reproduce all the non-verbal means of communication, such as gesturing. We have therefore developed a new method that learns plausible gestures from a body of motion capture data with simultaneously recorded audio (video). We can now synthesize plausible body language for avatars in real-time from speech. We plan to further improve this approach and include more advanced ways of interaction between the user and its avatar.
 |