Report on Specific Action 39-1: Cognitive Computer Vision Colloquium Prague January 12-13, 2004 http://cmp.felk.cvut.cz/cmp/events/colloquium-12-Jan-04/
Tomas Pajdla
The colloquium lasted two days and consisted of 11 oral presentations and 5 poster presentations. About 80 participants from the Czech Republic, Austria, Slovenia and the following speakers took part in the event:
Dmitry Chetverikov (SZTAKI Budapest, Hungary) Message for the Cognitive Vision Computer vision research reached sufficient maturity in image processing and invariant feature extraction (R.Deriche, M.Nielsen, J.Weickert, C.Schmid), geometry and optimization (R.Hartley, K.Daniilidis, S.-B. Kang, D.Martinec and B.Micusik), image matching (A.Zisserman, D.Lowe, A.Pinz, O.Chum, J.Kostkova), and object recognition and modeling (A.Shashua, L.Van Gool), to address problems of image and video interpretation on a higher level. There is clear shift of interest from the pure image formation understanding (camera modeling, multiple view geometry, appearance based modeling) to scene interpretation based on image matching. There is strong interest in using pattern recognition and machine learning techniques. It seems that a working approach to image matching has been found. Video indexing (similar to Google search engine) (A.Zisserman), automatic organization of image collections for mosaicing (D.Lowe), and stereo correspondence search for 3D scene reconstruction (D.Chetverikov) were presented. The approach works in real situations. While early image matching techniques relied on the epipolar constraint, the current approaches are more flexible. In many situations, they are able to provide matching that, albeit not always perfect, is good enough to give a reliable grouping of images into classes or extract features that occur repeatedly in images or videos. Cognitive vision, in its search for sensible and promising research goals (H.-H. Nagel), may now rely on the fact that current computer vision provides techniques for extracting stable or repeated features in images and maintaining their correspondences and relationships across changing views and lighting conditions. That poses the challenge to realize general goals of cognitive vision (http://www.ecvision.org/research_planning/Research_Dreams.htm) using currently available technology, i.e. to formulate an interesting 'cognitive vision task' and to solve it using the new image matching techniques. Tomas Pajdla
Center for Machine Perception
|
|
| Site generated on Friday, 06 January 2006
|