Section: New Results
Visual rendering
Real-time rendering for multiview autostereoscopic displays
Participants : Kadi Bouatouch [ contact ] , Christian Bouville, Rémi Cozot.
Multi-view auto-stereoscopic displays are now available at affordable cost and are getting widely used in virtual reality applications and 3D games. With their wide viewing zone, this type of display easily accommodates multiple viewers and does not require any head tracking. We have tested various virtual cameras settings with a view to keep the region of interest within the usable depth range of the display. We have also developed rendering methods allowing the generation of the interlaced multi-view image on the GPU. Our method accounts for the depth of field effect in case of multi-view auto-stereoscopic display. Real-time rendering on these displays poses a number of difficult problems: (1) real-time generation of several views of the same 3D scene, (2) choice of a particular sampling pattern of the displayed image requiring specific anti-aliasing procedures that results in a limitation of the usable depth range. We have already started to tackle these problems and hope to find out efficient solutions.
Bayesian Monte Carlo for rendering
Participants : Kadi Bouatouch [ contact ] , Christian Bouville, Jonathan Brouillat.
Global illumination techniques based on raytracing make intensive use of Monte Carlo techniques. The luminance incoming to the eye through a pixel is computed by performing Monte Carlo integration over a multi-dimensional complex function. Several methods have been proposed to reduce the variance of the result of Monte Carlo integration : stratification, importance sampling and control variates. We have proposed a global illumination method based on a new radiance cache approach [16] that does not need the use of any projection technique such as spherical harmonics or wavelets.
Bayesian Monte Carlo (BMC) techniques are widely used in the domain of Machine Learning, and relies on priors over the function of interest to improve Monte Carlo [13] computations. We have used BMC integration to speed-up the final gathering operation needed in global illumination computation. Choosing a prior for the incoming luminance function needs to be done carefully in order to effectively reduce variance. In addition, BMC integration requires inversion of large matrices, which can negate the advantages of casting less rays. We have proposed an adequate pre-computing scheme which does not introduce bias in the estimation. We have also proposed a more efficient method for determining optimal distributions of the samples, yielding then more accurate integration. We are currently working on different applications and extensions of BMC : aliasing, glossy surfaces, etc.
Real-time rendering for multiview autostereoscopic displays
Participant : Kadi Bouatouch [ contact ] .
Volumetric rendering is widely used to examine 3D scalar fields from CT/MRI scanners and numerical simulation data-sets. One key aspect of volumetric rendering is the ability to provide perceptual cues to aid in understanding structure contained in the data. While shading models that reproduce natural lighting conditions have been shown to better convey depth information and spatial relationships, they traditionally require considerable (pre)computation. We have proposed a shading model [28] for interactive direct volume rendering. It provides perceptual cues similar to those of ambient occlusion, for both solid and transparent surface-like features. An image space occlusion factor is derived from the radiative transport equation based on a specialized phase function. The method does not rely on any pre-computation and thus allows for interactive explorations of volumetric data sets via on-the-fly editing of the shading model parameters or (multi-dimensional) transfer functions while modifications to the volume via clipping planes are incorporated into the resulting occlusion-based shading.
