Daniel Meneveaux

Postdoctoral fellow in imager lab since october 1998 and untill october 1999.
With Alain Fournier

 
Research Interests :
 
• Computer graphics,
• Global illumination,
• Image based rendering.

Curriculum vitae (a very brief one...) :
 

• Postdoc, imager lab, Vancouver


- Pattern formation and growth in mammalian models,
 

I've been involved with Alain Fournier and Marcelo Walter on the integration of shape and pattern in mammalian models.

The aim was to integrate a pattern generation system which could effectively deliver a variety of patterns characteristics of mammalian coats, and a body growth and animation system that could use experimental growth data to produce individual bodies and their associated paterns totally automatically. We used the girafe to illustrate how the models can take us from a canonical embryo to a full adult girafe in a continuous way.


- Image based rendering.
 


 
• PhD thesis; Rennes, France

 

 

Topic :

Lighting simulation in complex architectural environments: Sequential and parallel approaches.

Introduction :

Nowadays, synthetic images are more and more used either for illustrating different kind of documents, or for applications needing the use of some precise images (medical imaging, movies, video games, driving simulation, etc.). With the help of techniques recently developed (augmented reality, lighting simulation for example), synthetic images become more and more realistic. Moreover, virtual reality allows a user to walk through an environment and even to interact with it. However, this kind of realism is still very demanding in terms of computing time, even with environments containing few objects. Obviously, the best realism is obtained by following physical rules defining the behavior of light. The simulation of such phenomenon implies the meshing of each surface within the environment resulting in a high number of small patches (called surface elements). In addition, several number of data structures are needed to accelerate computations which still remain time consuming. Though those techniques enable the lighting computation within small environments (one or two rooms within a building), they can not provide efficient results for very large buildings composed of several floors because the size of the  datastructures implied in the process is too high.

Partitioning :

The idea is to make possible this kind of computation by subdividing the problem into sub problems easier to solve. Our approach is based on the splitting of the environment into sub regions that we call cells. These cells are wanted to fit as best as possible to the topology of the scene. In other words, we desire each cell to correspond to one room of the building. With each cell is associated a list of portals (doors or windows for example). which allows us to establish visibility relationships between the cells. In other words, we determine for each cell the list of its visible cells, with the help of those portals. Finally, this partitioning results in a set of cells and a visibility graph. From the obtained datastructures, it is then possible to speed up the display of a complex environment and the lighting simulation within it. The idea is the following. A user standing in a given cell will only see the surfaces contained within this cell and the ones contained within its visible cells. We then only display the polygons belonging to those cells. For lighting simulation we use the same principle. A polygon belonging to a given cell C shoots its energy to the surfaces within C, and also to the surfaces of the cells visible from C.

One processor algorithm :

The lighting simulation algorithm that we divided works as the following. First, a cell
is chosen for emitting its patches (or mesh elements) energy. This cell is loaded into memory as well as all the cells visible from it and the radiosity computations are performed. Then, another cell is chosen, loaded into memory, etc. However, the ordering of choosing a cell is primordial in the lighting simulation process. For instance it is useless to unload a cell from memory if it is used in the next computation. In order to avoid this kind of problem, we propose 7 ordering strategies. Each strategy decides, according to its own criteria which cell reduces disk accesses or improves the convergence of the radiosity algorithm. Once lighting simulation has been performed, the resulting cells are saved onto the disk so that they can be used later for an interactive walkthrough.

Parallel Algorithm :

The results we obtained have shown a high difference of computation time between the 7 ordering algorithms. The best of those algorithms is the one that we called traveling salesman. Moreover, our lighting simulation algorithm seemed to be easy to be implemented in parallel. Consequently, we devised a new parallel approach using this technique, allowing the lighting simulation to be computed on several processors simultaneously. Processors can belong either to a parallel machine, a network of computers or both. They only have to be able to access a common disk. Each processor is then responsible for the computation of a set of cells. The obtained results are also stored into a set of files, in order to enable the user to interactively walk through the environment once the lighting simulation is computed.