Riveting Research

Click here to register for SIGGRAPH.

Click here

for a PDF.

SIGGRAPH 2003 Papers

Detailed Papers sessions tackle classic concerns including HumanBodies and Shadows.

IF YOU THINK IN STATISTICS, CONSIDER THIS: The Papers Committee ofSIGGRAPH 2003 has been 18% busier than last year. According to Paperschair Jessica Hodgins, associate professor at the School of ComputerScience at Pittsburgh's Carnegie Mellon University, the 2003 Committeeobtained a record number of submissions at 424 and accepted 81 papers;a 19% acceptance rate similar to that of past years.

But ultimately, numbers are not the point, Hodgins says, explainingthat the Papers Committee continued the tradition of not countingpapers, and instead accepted every paper they thought should appear atSIGGRAPH. Although this resulted in some overlapping sessions, it alsoresulted in the "best possible overall program."

The acceptance process began early this year, when the five-memberAdvisory Committee met for a Papers Sort meeting, during which eachpaper was assigned to two committee members. Those members eachassigned one or two outside referees to review the papers as well.These initial reviews were submitted to the authors, and during a newlyadded Rebuttals phase of the review process, the authors were allowedto correct factual errors via posts to an electronic bulletin boardsystem. Approximately 75% of the authors took advantage of theRebuttals phase. "Despite the extra workload, the committee felt thatthis extra step in the review process improved the quality of the finaldecisions," Hodgins says.

Once the papers were conditionally accepted in late March, they wentthrough a second review process, during which the authors responded torequired changes and made additional changes as appropriate. The paperswere read once more, and the final versions were approved.

Graphcut Textures and Wang Tiles

Overall, the 2003 papers represent significant leaps forward in manyareas. One such area is texturing. In "Graphcut Textures: Image andVideo Synthesis Using Graph Cuts," a paper in the "Texture Synthesis byExample" session, authors Vivek Kwatra, Arno Schödl, Irfan Essa,Greg Turk, and Aaron Bobick of the Georgia Institute of Technology/GVUCenter introduce a new algorithm for image and video texturesynthesis.

In their approach, as stated in their abstract, "patch regions froma sample image or video are transformed and copied to the output andthen stitched together along optimal seams to generate a new (andtypically larger) output."

According to the authors, this process is different from currentapproaches in that instead of choosing the size of the patch "apriori," a graphcut technique determines the optimal patch region forany given offset between the input and output texture. In their paper,the authors explore graphcut textures in both 2D and 3D to performvideo texture synthesis and traditional image synthesis. They also showresults for synthesizing regular, random, and natural images andvideos, and they demonstrate how the graphcut technique can be used tointeractively merge different images to generate new scenes.

Another paper in the "Texture Synthesis by Example" session is "WangTiles for Image and Texture Generation," authored by Michael F. Cohen(Microsoft Research), Jonathan Shade (Wild Tangent), and Stefan Hillerand Oliver Deussen (Dresden University of Technology). In this paper,the authors present a new stochastic algorithm for non-periodicallytiling a plane with a small set of Wang Tiles. Wang Tiles are squaresin which each edge of each tile is assigned a color. The authorsdiscuss how Wang Tiles can be used and re-used to efficiently createlarge expanses of complex textures, patterns, or prelighted geometry atruntime, while avoiding obvious visual artifacts of repetition.

Reanimating the Dead and Skinning the Animated

Another area of note is human modeling and animation. The "HumanBodies" session features five papers, each highlighting a breakthroughin this arena.

In "Reanimating the Dead: Reconstruction of Expressive Faces fromSkull Data," for instance, authors Kolja Kähler, Jörg Haber,and Hans-Peter Seidel of MPI Informatik explore the possibilities foranatomically correct modeling and physics-based simulation in aninteractive environment.

They do this using Medusa, a facial modeling and animation systemthat Haber and Kähler developed for research and educationalpurposes. As the authors explain in their abstract, the system includestools and techniques for a variety of subtasks that must beaccomplished to generate photorealistic facial animations in realtimeon standard PCs. In their paper they discuss the topics of dataacquisition, muscle modeling, skin simulation, textures and rendering,multiresolution surfaces, speech synchronization, and anthropometricmodeling.

In "Building Efficient, Accurate Character Skins from Examples,"authors Alex Mohr and Michael Gleicher from the University ofWisconsin, Madison, present an automated framework for buildingconvincing skin deformations for interactive applications. As Mohr andGleicher explain, good character animation requires subtle yet detaileddeformation. Commercial animation packages provide tools for achievingsuch deformation; however, these programs are not optimal forinteractive applications because they are too slow to compute orrequire too much memory.

The method outlined in this paper begins with an arbitrarily riggedcharacter in an animation system. The authors export a set of examplesconsisting of skeleton configurations paired with the deformed geometryas static meshes, and then use these examples to fit the parameters ofa deformation model that best approximates the original data.

According to the abstract, "the underlying deformation model we useis an extension of the traditional linear blend skinning model widelyused in interactive applications. This extension allows us to capturesubtle and detailed effects required for believable characters butremains nearly as fast and compact as linear blend skinning."

Skin deformation also is the focus of "Continuous Capture of SkinDeformation," presented by Peter Sand and Jovan Popovic from theMassachusetts Institute of Technology, and Leonard McMillan from MITand the University of North Carolina at Chapel Hill. In this paper, theauthors describe a method for acquiring deformable human geometry fromsilhouettes.

"Our technique uses a commercial tracking system to determine themotion of the skeleton, then estimates geometry for each bone usingconstraints provided by the silhouettes from one or more cameras,"write the authors. "These silhouettes do not give a completecharacterization of the geometry for a particular point in time, butwhen the subject moves, many observations of the same local geometriesallow the construction of a complete model."

The authors' reconstruction algorithm provides a simple mechanismfor solving the problems of view aggregation, occlusion handling, holefilling, noise removal, and deformation modeling.

Mocap and Range Scans

The paper "Free Viewpoint Video of Human Actors," by Joel Carranza,Christian Theobalt, Marcus A. Magnor, and Hans-Peter Seidel of MPIInformatik, describes a system that uses multi-view synchronized videofootage of an actor's performance to estimate motion parameters and tointeractively re-render the actor's appearance from any viewpoint.According to the abstract, "the actor's silhouettes are extracted fromsynchronized video frames via background segmentation and then used todetermine a sequence of poses for a 3D human body model. By employingmulti-view texturing during rendering, time-dependent changes in thebody surface are reproduced in high detail."

Among the Fluids and Smoke papers are "Animated Supsended ParticleExplosions" (pictured), and a groundbreaking presentation from ILM onsmoke simulation.

The authors write that the motion-capture subsystem, which runsoffline, is non-intrusive yet yields robust motion parameter estimates,and the rendering system, which runs at realtime frame rates, yieldshighly natural results.

The final paper in the "Human Bodies" session is "The Space of HumanBody Shapes: Reconstruction and Parameterization from Range Scans."Authored by Brett Allen, Brian Curless, and Zoran Popovic of theUniversity of Washington, this paper describes a method the authorsdeveloped for creating a whole-body morphable model based on 3D scannedexamples.

Shadows and GPU Computation

One subject that has been covered only periodically in the SIGGRAPHPapers Sessions over the last few years, but this year has an entiresession devoted to it, is shadows. The appropriately titled "Shadows"session features four papers that describe the latest advances inshadow generation. Indeed, this session covers everything fromextracting shadows from one natural scene and inserting them intoanother, described in "Shadow Matting and Compositing," to a newalgorithm for interactive generation of hard-edged, umbral shadows incomplex environments with a moving light source, outlined in"Interactive Shadow Generation in Complex Environments."

Also being presented are papers on meshing and surfaces,measurements for rendering, character animation, and animation ofsmoke, explosions, and cloth. This year also highlights the subject ofalgorithms for GPUs. Hodgins says that GPU computation, along with thecapture of motion and body shape from video, are two techniques thatmay become commonplace in the not-too-distant future.

This year's Papers Sessions truly represent the best academic andindustry research in computer graphics. "It's an exciting program,"Hodgins concludes. "We're presenting a lot of new ideas and newtechnologies that will have a large, positive impact on the growingfield of computer graphics."

Papers can be purchased as part of the SIGGRAPH 2003 technicalmaterials, which include the Full-Conference DVD and "ACM Transactionson Graphics." The DVD contains the electronic version of the technicalpapers, images, and supplemental material; the course and tutorialnotes, including supplemental material; and the permanent record of theEducators Program, Emerging Technologies, Sketches & Applications,Special Sessions, and Web Graphics programs, as well as the Art Galleryand Computer Animation Festival. "ACM Transactions on Graphics"contains the SIGGRAPH 2003 technical papers and the ACM SIGGRAPHawards. Publications are available for purchase at www.siggraph.org/publicationsand are available to ACM SIGGRAPH members at substantial discounts.

SIGGRAPH 2003 Papers Sessions

Sunday, July 27

Fast-Forward Papers Preview
6:00-7:30 p.m.

Monday, July 28

Session: "Texture Synthesis by Example"
10:30 a.m.-12:15 p.m.

Session: "Images, Video, and Texture"
1:45-3:30 p.m.

Session: "Parameterization"
3:45-5:30 p.m.

Session: "Precomputed Radiance Transfer"
5:45-7 p.m.

Tuesday, July 29

Session: "Character Animation"
8-10:15 a.m.

Session: "Visualization and Printing"
10:30 a.m.-12:15 p.m.

Session: "Surfaces"
1:45-3:30 p.m.

Session: "Shadows"
3:45-5:30 p.m.

Session: "Perception and Manipulation"
5:45-7 p.m.

Wednesday, July 30

Session: "Human Bodies"
8-10:15 a.m.

Session: "Light Fields and Visibility"
10:30 a.m.-12:15 p.m.

Session: "Points"
1:45-3:30 p.m.

Session: "Modeling and Simplification"
3:45-5:30 p.m.

Session: "Reprise of UIST '02 and I3D '03"
3:45-5:30 p.m.

Thursday, July 31

Session: "Fluids and Smoke"
8-10:15 a.m.

Session: "Scattering and Reflectance Measurement"
8-10:15 a.m.

Session: "Hardware and Displays"
10:30 a.m.-12:15 p.m.

Session: "Design and Depiction"
10:30 a.m.-12:15 p.m.

Session: "Dynamics"
1:45-3:30 p.m.

Session: "Computation on GPUs"
3:45-5:30 p.m.

Session: "Meshes"
3:45-5:30 p.m.

SIGGRAPH 2003 Papers Advisory Board

CHAIR: Jessica Hodgins, Carnegie Mellon University

Joe Marks, Mitsubishi Electric Research Laboratories (MERL)

Peter Shirley, University of Utah

Greg Turk, Georgia Institute of Technology/GVU Center

Marc Levoy, Stanford University

Click here to register for SIGGRAPH.