X-M1 & M2R project :
« Real-time walk-through the Milky-Way:
continuum to points, on-the-fly generation,
and mixing with catalog »


Fabrice NEYRET   - Maverick team, LJK, at INRIA-Montbonnot (Grenoble)


One of the Graals of Computer Graphics is the realistic rendering of ultra large and detailed scenes, possibly with real-time walk-through. In such case we can’t explicitly render or even generate and store everything once. Scalability and on-the-fly adaptiveness is thus required: we need to produce and process only what is currently visible, and directly at the relevant resolution.
Virtual exploration of galaxies (see some of our previous work) is a challenging example of it : we can’t generate all stars, even if we could we still couldn’t render all of them in realtime, and when the density is so high that they appear as a fluid it would be a waste to generate individual points. Also, we can’t simply sample and place random points in the Universe: instead, we need to find directly what falls in any given local region. Moreover, we need need to mix known data (catalog of starts that are very bright or close to Earth) with stochastic generation when only the statistics is known.
( Note that the same problem occurs in other contexts, e.g. tree generation in games or simulators, where some special landmark trees have to mix with heterogeneous statistics. )

Description of the subject

Uniform probability for coordinates turns into Poisson law when looking for probability of space occupancy (i.e., Poisson distribution). Since this law has fractal properties, and converges towards Normal law, continuum emerging when the variance get weak (i.e., law of large numbers), we want to explore top-down strategies to generate the content needed for a given point of view, either continuous or discreet, and withholding local known data out of stochastic generation. And since modern graphics cards allow massive computations, we want to rely on data structures and algorithms directly running on the GPU (as for our GigaVoxels or Proland projects).