* The objective of the project/research hypothesis The objective of proposed project is to develop an original approach for creation of models representing dynamic, time-varying physical scenes observed with cameras or 3D scanners. Proposed representation will allow for precise reconstruction of a digitalized scene for visualisation or reproduction, processing of changes, such as motion or shape change of objects and robust segmentation. * The Method The approach proposes to use original approach, called the Atomic Shape Network. It is a special case of a Dynamic Bayesian Network, where variables represent: scene structure (as a graph of elementary shapes), observations (depth and spectral information from cameras or scanners), and scene properties (e.g. motion of objects). Subsequent research task in the project correspond to theoretical problems associated with such representation -- among others, with estimation of parameters of atomic shapes, learning the network structure, detecting and encoding changes in scene, incorporation of spectral information in shape description. * Importance and influence of the results on science, civilization, society The result of the project will be an original probabilistic framework, tools and methods for robust building and on-line updating of 3D scene representation. The internal structure of the scene model will allow both detailed reproduction and access to higher level information (parts of objects, motion segmentation) easily available for machine learning algorithms (recognition, indexing). Problems of reconstruction and representation of 3D scene are a central element of the discipline of processing of 3D information, both in the acquisition and storage, and for most domains of applications. Thus the results of this project will have a potential impact for a number of disciplines using the scene reconstruction, e.g. robotics, computer graphics, human-computer interaction.Importance and influence of the results on science, civilization, society} The result of the project will be an original probabilistic framework, tools and methods for robust building and on-line updating of 3D scene representation. The internal structure of the scene model will allow both detailed reproduction and access to higher level information (parts of objects, motion segmentation) easily available for machine learning algorithms (recognition, indexing). Problems of reconstruction and representation of 3D scene are a central element of the discipline of processing of 3D information, both in the acquisition and storage, and for most domains of applications. Thus the results of this project will have a potential impact for a number of disciplines using the scene reconstruction, e.g. robotics, computer graphics, human-computer interaction.