任重 图形学 绘制 图形处理单元 真实感 软影Associate Professor
Before returning to Zhejiang University to co-found the Graphics and Parallel Systems (GAPS) Lab with Prof. ZHOU Kun in Apr. 2010, I spent a happy five years in the Internet Graphics Group, Microsoft Research Asia, first as an intern then as an associate researcher. I received my Ph.D degree in computer science from Zhejiang University in March, 2007. I also hold a Master's degree in mechanical engineering and a Bachelor's degree in information engineering, both from Zhejiang University. My research interests include real-time rendering of soft shadows and participating media, spherical harmonics for real-time rendering, and real-time global illumination on GPUs. My Ph.D thesis (in Chinese, with English abstract) is about precomputation-based methods in interactive global illumination. During the past several years, I've been really fortunate to work with my collaborators on the interesting graphics projects below. That's really a lot of fun.
In my spare time, I like Chinese calligraphy, music, and swimming.
- IPC member: Pacific Graphics 2010, 2013
- reviewer: ACM SIGGRAPH 2012-2013, ACM SIGGRAPH Asia 2008/2010-2013, IEEE Vis 2010, PG 2009-2012, EG 2009,2014, HPG 2013, ESPGV 2011, Euro-Par 2013, TVCG, TOG, CGF, C&G, NSFC, JCAD.
- using cone tracing for hair/fur
- approximate away in-cone variation of shading, opacity and occlusion of samples
- form reflection/refraction cones using ray differential
- significant reduce rendering the time needed for generating noise free results due to reduced required supersampling rate
- DOF, motion blur, reflection/refraction of fur/hair are natually supported
- easy to integrate into existing ray tracing framework
- not suitable for rendering of thick fur strands
- large cones bring down performance, as they may intersect large number of strands
- motion blur support is preliminary
downloadable:
- depict the boundaries of visible regions using piecewise linear functions
- convert the shading computation into double product integrals
- approximate the double product using Legendre polynomials locally in visible regions
- improve the shadow accuracy of existing real-time all-freq PRT method
- triple product decomposed to a number of double products in the visible regions
- on-the-fly extraction of visible regions from spherical signed distance functions (SSDFs)
- limited ability to handle complex visibility due to interpolations of SSDFs
- currently no support for dynamic scenes
downloadable:
- process multiple frames of a pixel in parallel to maximize exploitable coherence
- reduce required trace resampling rate using a parallel analytical visibility algorithm
- precompute and store the intersections of view rays and the moving object to improve the runtime performance
- programmable motion blur made more practical
- improved performance for rendering programmable motion blur for an order of magnitude
- scalable to large scenes consist of hundreds of moving characters
- mesh connectivity requires to be constant
- currently no support for shadows, reflections and refractions
downloadable:
- a more accurate and efficient particle-based algorithm to simulate bubble dynamics and interactions
- an approximation of foam geometry, by treating bubble particles as the sites of a weighted Voronoi diagram
- a set of bubble interaction forces to handle miscellaneous foam behaviors
- explicitly address the volume loss issue in foam simulation
- accurately model various interaction effects among bubbles
- does not handle bubble surface deformation, therefore not suitable for large bubbles
- nonlinear interaction forces in the real world cannot be modelled
downloadable:
- discretize diffusion equation based on discrete divergences and derive analytic formulas can be derived for DE matrix elements
- multi-resolution GPU DE solver that adapts to the frequent changes in topological structure of dynamic objects
- enables interactive rendering of translucent cutouts
- allows interactive topology changes of translucent object
- similar accuracy as the FEM solver based on Guassian quadrature integration
- per-vertex shading, representable details limited by mesh resolution
- controlling of tetra aspect ration is required durring the cut
downloadable:
- introduce the analytic form of non-linear rays in refractive media with constant refractive index gradient, based on which a ray tracing algorithm is developed;
- extend the algorithm to general non-constant refractive media through tetrahedra-based representation or voxel-based representation;
- derive the analytic form of non-linear rays in mirages.
- analytic forms of rays and fold points;
- larger tracing steps, leading to superior performance.
- no simple analytic form for ray-scene intersections, need to resort to numeric methods;
- piece-wise linear approxiation of refractive index function, may require impractical large number of tetra for very complex functions.
downloadable:
- decouple light integration complexity from light transport complexity;
- model environment light with a set of SRBFs;
- address light integration complexity by precomputation;
- handle self shadowing by convolution optical depth map;
- handle mutiple scattering by dual scattering and sparse sampling and interpolation.
- the first interactive hair rendering algorithm under environment lighting;
- support light animation;
- support dynamic hair volume.
- precomputation per scattering function;
- ignore eccentricity in the scattering function.
downloadable:
- tabulate per-vertex transfer from global incident directions to local incident directions;
- use biclutering to directly compress the transfer matrix.
- all-frequency biscale transfer, support highly-detailed surface under env-lighting;
- high compression ratio with no quality lost;
- operates directly on pixel basis, fully dynamic env-lighting can be directly supported.
- static scene geometry;
- rather costly pre-computation(~1 hour).
downloadable:
- move all stages of basic Reyes pipeline to GPU;
- introduce a dynamic stage scheduling algorithm to maximize the available parallelism at each individual stage of the pipeline while keeping data in GPU memory;
- minimize inter-GPU communication by a multi-GPU scheduling algorithm based on work stealing, which can be easily combined with the stage scheduling algorithm above;
- implemented with BSGP.
- could be the fastest Reyes renderer on PCs;
- support multi-GPU and out-of-core texture;
- quality is comparable with PRMan.
- system will fail if the grids and their bounding boxes do not fit into the vram;
- currently only support basic Reyes features. no ray tracing, no photon mapping...(coming soon)
downloadable:
source code, test scenes
- accurately evaluate source radiance and its gradient at a relative small number of sample points in the volume, and use gradient-based interpolation to approximate the source radiance elsewhere;
- sample the interpolation error in the valid sphere of each sample, and dynamically splitting those with higher estimated error on a hierarchical grid structure.
- use source radiance splatting to construct the source radiance volume, followed by a ray march to yield the radiance to the view.
- all-frequency lighting effects such as light shaft, volumetric shadows and inhomogeneous glowing;
- editable media property;
- good scalability with regard to volume resolution and number of lights;
- precomputation can be avoided for analytical medium representations.
- single-scattering only (though it's possible to be combined with the technique below to incorporate multiple-scattering);
- constant phase function.
downloadable:
- as a preprocess, decompose density field into an analytical low-freq RBF representation and a residual field which can be compressed by perfect spatial hashing;
- evaluate source radiance at RBF centers, taking into account only the RBF representation;
- use SHEXP to accelerate single scattering computation, and BSGP to accelerate multiple scattering;
- compensate for the residuals in the final ray marching pass.
- fast smoke rendering with single/multiple scattering effects and fine details;
- editable smoke property;
- moderate preprocessing time, less than an hour for a 600-frame animation.
- low-frequency lighting, can't handle strongly directional phase functions;
- can't handle real dynamic medium due to the need for precomputation.
downloadable:
- approximate block geometry with sphere blockers;
- accumulate tabulated sphere SH visibilities, in log space;
- approximate SH exponentiation by an optimal linear model + DC isolation & scaling/squaring techniques;
- use sphere hierarchy + receiver clustering to accelerate run-time computation.
- fast rendering of soft shadows of skin-animated characters;
- very small precomputation cost: the tabulation of the log visibility of a unit sphere is reused for all sphere blockers, and the sphere approximation for a rest pose of a character is skin-animated with the geometry.
- low-frequency lighting, very soft shadows;
- can't support generic local lighting.
- per-vertex shading, shadow quality and performance depend on model triangulation. (an extension to per-pixel shading by Peter-Pike Sloan et al. is available)
downloadable:
preprint, video(49.7MB), slides(31.9MB), supplements(3.1MB), test data#1(tumbling man)(8.9MB), test data#2(walking man)(5.2MB), test data#3(slipped walking man)(1.5MB), test data format, bibtex
- tabulate the intersection of scene geometries in a 4D line space to form i-fields;
- use visibility query & splatting in parameter space to accelerate run-time computation;
- use photon splatting in parameter space to compute indirect bounces.
- visibility query made fast;
- fast scan-line conversion for i-field generation;
- reuse i-field for object instancing and object transformation.
- no compression, scalability is low;
- limited shadowing feature reprensentability due to line space resolution.
downloadable:
downloadable: