K.U.Leuven > ESAT > PSI > Visics > Research > Topics > Item 1.11

Efficient representation of 3D scenes

L. Van Gool, L. Van Eycken, B. Deknuydt

In cooperation with IMEC DESICS-MICS, research is done on efficient representation of 3D scenes, both with respect to transmission and rendering efficiency. Most effort is been put on time varying aspects of mesh and to a lesser extent, texture. This is needed, because advanced 3D capture systems are now able to deliver highly detailed models of time varying objects at up to 25Hz, generating huge amounts of data. Only specialized data compression techniques will be able to reduce this significantly.

A first approach consists of exploiting time correlation between successive meshes. PSI developed a motion clusterer, which represents complex motion of e.g. a human face by subdividing it into patches with each a generalized motion (translation, rotation, expansion/contraction). The image shows regions of consistent motion on a talking human face.

Another path is to try and exploit the combined effect of mesh and texture. Complex texture is known to make the human observer less sensitive to errors in geometry. As we know the texture, we can drive mesh simplification algorithms more intelligently.

The longer-term target is to obtain efficient modelling of objects with a complex surface structure (think e.g. of a human face with hair). This is a tempting task as either the quality is not convincing, or the model size (and the effort to create it manually) is prohibitive. A conceptually new, scalable approach has been developed, modelling 'smooth' objects with geometry alone, 'rougher' objects with additional bump maps and 'very rough' objects with viewpoint dependent generated textures.