Skip to main content

Dynamic Minimal prior knowledge for model based Computer Vision and Scene Analysis

Final Report Summary - DYNAMIC MINVIP (Dynamic Minimal prior knowledge for model based Computer Vision and Scene Analysis)

Efficient solutions for open problems in computer vision are often achieved with the help of suitable prior knowledge, e.g. stemming from labeled databases, physical simulation or geometric invariances. Especially for scene analysis, database knowledge can become so large and complex, that it cannot be integrated efficiently for optimization. On the other hand, there exist geometric priors which are efficient and compact, but they have to be integrated and exploited explicitly in vision systems. As a consequence there is need to develop methods to conclude from (statistical) database knowledge to geometric prior knowledge and therefore to achieve compressed priors which contain the relevant information from a given database.

Advancing prior knowledge means to seek for the essence and granularity of priors. My work focussed on the derivation of geometric priors from statistical information in the context of human motion understanding.
In several papers we documented on a human shape database, human motion database and a human shape/motion database. We explored factorization methods to integrate priors and investigated physical models.