Research On Fluid Surface Reconstruction

Animating fluid surface is becoming more and more popular in the field ofboth computer games and movies. The rich details of water have been widelyused in the CG, virtual reality, game, image design and etc. In movie making,there are trends that more and more movies contain scenes of huge water flows.On the other hand, with the help of the IMAX technology, the resolution of themovie is increasing dramatically, which require more surface details andqualities when the audience watch the movie so as to convince them that thewater scene is real.Traditional handmade water scenes are impossible in today’s newrequirements. Therefore, physically based water simulation and realistic fluidsurface rendering method has become a very hot research field of computergraphics. Physically based water simulation is proceeding according to thefamous Navier-Stokes equations. To solve Navier-Stokes equations in thecomputer, a discrete solving method must be proposed. The very well-knownEuler grid methods and Lagrange particle methods are the most common solverfor the Navier-Stokes equations. While the Euler grid methods have higheraccuracy and maintain incompressible properties of fluids, Lagrange particlemethods have low numerical viscosity，far more details and runs on the fly.In our paper, we start from the SPH method which is a form of Lagrangeparticle methods, generating huge amount of space particles which carry themass of the fluid. To generate surfaces from these space particles, level setmethod, explicit surface flow and implicit methods like density filed and distancefield have been proposed.Comparing to previous works, our work proposes a new distance fieldcalculation method. Unlike previous methods which sample the particles directlyin the grid point, our method uses the geometry median to define the samplingcenter. By doing this step, the error and inaccuracy caused by outliers and minorsmall scale surfaces.Besides, we proposed a new distance field model that focuses on the shapeand position of the sample region. While the previous method uses isotropy sampling region to pick the surface particles, our method applies spacetranslation on the sampling kernel so as to let the sampling region stretchedalong the surface and shrunk along the surface normal. The strategy caneliminate the problem causing by the irregular placement of the SPH particleposition.Finally, we use PCA to generate a local plane from the particle position. Thedistance function is defined using the distance from the grid point to that surfaceplane, which results a more accurate distance implicit function.