 | | Home | | Research | | Teaching | | Forums | | Staff | | Contact | | Search | | Intranet | |
|
||||||||||||||||||
|
||||||||||||||||||
|
||||||||||||||||||
|
||||||||||||||||||
|
|
|
|
|
|||
|
Research Projects Student and Thesis
Projects, Jobs Scientific
Publications |
Image and Video Analysis & Synthesis ProjectsOur group performs research on image and video analysis & synthesis with a strong focus on stereo processing, image matting, video coding, and visualization. We use our stereo-derived 3D models to compute non-photorealistic visualizations in imitation of artistic drawings or paintings produced by hand. In collaboration with industrial partners, we seek solutions to 3D video/TV applications and explore the optimized implementation of our algorithms on chip.
Stereo AnalysisIn a stereo matching approach, two images are recorded from two slightly displaced cameras. This images are overlayed to infer depth information. The task of overlaying the images is known as stereo matching problem. Being able to solve the stereo matching problem is important for numerous applications in computer vision. Without being exhaustive, application examples include automated extraction of three-dimensional models (terrain and city models, e-commerce, cultural heritage), robot navigation (obstacle avoidance, localization), three-dimensional tracking (surveillance, pose estimation, human-computer interaction), depth segmentation (z-keying), industrial applications (quality assurance) and novel view generation (free viewpoint video), to name just a few of them. The work on stereo matching at the IMS concentrates on two factors that currently represent a limiting factor for the success of stereo vision-based applications. Firstly, our major focus lies on improving the quality of stereo matching results. Currently, the results of standard techniques are still often too poor for practical applications. Secondly, we work on the development of computationally fast methods. Even a perfect stereo matching result does not help for a large number of applications, if its computation took a couple of hours. Image and Video MattingImage/video segmentation and matting are fundamentally important operations in many image/video editing and compositing tasks, with promising applications in the entertainment industry. Matting refers to the process of extracting arbitrarily shaped foreground objects from images or videos by correctly recovering the transparencies of the object. On the other hand, the seamless insertion of an extracted object into a new scene is known as compositing. Matting is a severely ill-posed problem and therefore user interaction is commonly used to make the problem well posed. A very common form of user interaction is the trimap interface, where the user marks those regions in an input image that obviously belong to the foreground object and background, respectively. Transparency values are then computed for the remaining unspecified region. Other algorithms are also capable of working on sparse trimap input which is denoted as scribbles. VisualizationAn important part of our research activities is dedicated to the development of visualization techniques, with a focus on generating non-photorealistic views from real images and novel visualization methods in the context of 3D video/TV. Our research on non-photorealistic rendering deals with the automated generation of artistic drawings and paintings from real images as input. Some examples of artistic drawings, sketches, and cartoon-like representations produced by our algorithms are shown below. We have also developed techniques for non-photorealistic stereoscopic rendering. For more information, please, take a look at these two dissertations:
Video CodingVideo coding deals with the representation of video data for analogue and digital video signals. The focus of our group is on the digital side where we already have a discrete version of a video signal. We typically record this signal with a digital camera and get a minimally processed version of the image sensor data (i.e. raw data). After the raw video data is available, video coding techniques typically use various forms of data compression and reduction. This is normally necessary since storing and transmitting of raw video data typically exceeds the available storage size and bandwidth. A wide range of data compression techniques (e.g. spatial and temporal prediction, context-adaptive entropy coding) and data reduction techniques (e.g. chroma sub-sampling, quantisation based techniques) exist in the field of video coding. Next to compression and reduction, other aspects have to be addressed by state-of-the-art video coding techniques:
One focus of our work on video coding at the IMS lies on the parallelization of video coding algorithms. The high computational complexity and the memory transfers in single core video coding systems currently represent a limiting factor for high resolution videos such as 1080p60. We focus on finding efficient splitting approaches of the major decoding and encoding algorithms and on the design process of multi-core video coding architectures. |
|||||
|
Print
this Page |
|
