Computer Vision Lab

Computer Vision aims at the extraction of information from images for steering and controlling machines (e.g. robots) or for steering and supporting human action and decision-making. The data can be still images, image sequences (video), volume sequences (e.g. MRI) or even 4D sequences (volume and time). Images can stem from very different physical sources (visible light, acoustic, radar, infrared, magnetic resonance, etc.).

Images are usually very high dimensional. To be able to handle this high dimensionality an important objective of computer vision is to reduce the dimensionality by low-level feature extraction (key points, edges, lines, texture, variance, etc.). A next step is to use these features to segment the images into useful objects, which can then be analyzed, measured or interpreted. The focus in our computer vision lab is on segmentation and analysis of multidimensional data (image sequences, multiple cameras, 3D/4D medical data like MRI and CT). Segmentation has to be seen in a broad sense: distinguishing the important image information from the non-important image information. Analysis is mostly video analysis: correspondence estimation (motion, disparity), object/model detection, object/model tracking, etc. The main research areas currently covered are 3D imaging (camera calibration, disparity estimation, 3D reconstruction 3DTV/Free viewpoint rendering), biomedical imaging (medical image segmentation, 2D/3D model reconstruction), social or human signal processing (pose and gesture recognition and tracking) and surveillance (video object detection, recognition and tracking).

A major application of computer vision in our group is the 3D reconstruction of objects from multiple cameras or moving camera(s) for e.g. face recognition, remote handling, 3D editing or 3D teleconferencing. An important issue for this application is the calibration and registration of multiple cameras. This includes modeling the image formation and projection process and determination of the geometric properties and position information of the sensors. Research items are high accuracy calibration of 3 or more cameras and calibration and registration of multiple source cameras. A related topic is correspondence estimation, that is which pixel in an image corresponds to which pixel in another image. This can be images from the same cameras (motion estimation) or different cameras (disparity estimation). The correspondence information can then be used for e.g. depth estimation through triangulation, for data compression (coding) or view reconstruction from arbitrary positions for e.g. immersive teleconference systems. For biomedical image processing and analysis we collaborate with Leiden Medical Centre (LKEB, Radiology lab) and with Erasmus Medical Centre (BIGR, BME) in Rotterdam Some recent projects are described in the section Research Projects.