Section: New Results
Participants : François Sillion, Gilles Debunne, Xavier Décoret, Denis Haumont, Elmar Eisemann, Samuel Hornus.
Fast Scene Voxelization and Applications
In this approach standard graphics hardware is used to dynamically calculate a voxelized representation of the scene, optimized for current cards and and involving minimal data transfer. Voxels are a simplified representation of the scene in form of a constant number of information with respect to a chosen resolution.
Until now, the creation of such a structure remained expensive. Using our approach 1,000,000 polygons can be voxelized in the order of milliseconds.
Several applications arise; like attenuation by semi-transparent materials, volume estimation for refraction and scattering and fast, high quality shadow calculation (culling / clamping for shadow volumes).
The technique is especially promising, as the announced features of the next generation graphics hardware will be highly adapted to our approach. The paper has been accepted to I3D 2006  . It has been selected for re-publication as a sketch at Siggraph  .
Plausible Image Based Soft Shadows Using Occlusion Textures
Computing soft shadows in real-time is a challenging problem. Computations are inherently complex. However, the human brain is relatively bad at ``analysing'' soft shadows, and approximate shadows are easily accepted as realistic. Based on that observation, we can tradeoff some accuracy for some computation speed. We proposed a method for computing ``plausible'' soft shadows on the GPU. The geometry of the scene is approximated by a set of planar bitmasks (slices), using the fast scene voxelisation introduced in previous section. We then use the NBuffer recently introduced by Xavier Décoret to pre-compute convolution of this bitmasks by different kernel size. As shown by Cyril Soler in an older work, at a given point, the soft shadows caused by one slice is given by the convolution of the slice with a kernel whose size and location depends of the point and the light source. Combining our GPU based encoding of slices and convolutions, we are able to compute this result very efficiently in a fragment shader. The problem that remains is the combination of the shadows caused by each slice. We introduce a novel scheme, based on probabilities, that perform significantly better than previous methods, although it is not exact. As a result, we can compute very appealing soft shadows (see Figure 11 ) on arbitrary scenes (complex geometry, high polygon count, animated scenes) very fast, with all computations taking place on GPU. The result have been published at Sibgrapi  and has been selected for re-publication as a journal paper at CGF (acceptance is on its way).