VR 2003 report

This year's IEEE Virtual Reality conference was held in the Biltmore hotel in Los Angeles. I went there for a tutorial and three days of regular conference business and was very happy to have done so. In the following you will find a report on what impressed me most.

Saturday: Image-based rendering tutorial

Martin Jägersand (U Alberta), Zachary Dodds (Harvey Mudd College), Dana Cobzas (U Alberta), Darius Burschka (Johns Hopkins), and Keith Yerex (Virtual Universe Corp.) gave a very interesting tutorial on image based rendering. A review of the basics, in particular the various view models, was followed by Martin's research on dynamic textures. Dynamic textures are parameterized by the direction they are view from. This function is learned from relatively extensive sampling of the view sphere with SVD methods, but it produces very nice view interpolation results. Even better, the dynamic textures can also be parameterized by non-rigid objects so that for example the flexion of an arm looks fairly natural including the cloth movements. More information is available at the VR site. I also have the tutorial script and many notes.

Monday

The actual conference started with Steven Feiner's keynote address on outdoor VR. It was an interesting and inspiring talk, but I knew most of it from Tobias' presentations. I later had a brief personal conversation with Prof. Feiner about user interfaces for AR, and he mentioned a toy data glove for $80 - 6 DOF tracking plus finger flexion, that's pretty much unbeatable!

Next I checked out the exhibition hall. Most noticable was Vicon's stand because it had the most equipment and was mostly empty. Too bad. Andy Beall's team of course had the most fun demo - chase ducks with bouncing balls! The entire floor was filled with thumbing sounds and quacking ducks... InterSense showed a great new tracker development, fiducial tracking combined with inertia tracking to cover large spaces. Looked very robust, accurate, and fast, and will hopefully be less expensive than other wide area trackers. Eric Foxlin also presented a paper about this new development, see Wednesday. Sebastien Grange had a stand demonstrating two versions of the Force Dimension haptic force feedback devices. They have a very natural feel to them and are very rigid in their response.

A panel discussion after lunch discussed VR applications in industry. Randy Smith from GM concluded after more than a decade of using VR for design studies that people experience virtual cars (especially reaching distances) not significantly different from real cars. So that's good, VR does the job much cheaper than carving models from clay. To the question "besides faster, better, cheaper - what does the car industry want from VR?" he pointed into the direction of perceptual validation. That is, how do people perceive virtual objects, shapes, sizes, distances, etc. vs. their real counterparts? The other three panel members did in my opinion not present any novel thoughts about VR.

I want to mention two papers, one on calibrating a head mounted projective display for AR purposes. It's a novel idea, but I wonder about its applications. Ulrich Neumann presented Augmented Virtual Environments (AVEs), where live video was overlayed over virtual structures. Neat was how they constructed the VR model (from LiDAR data, overflight 3D data): the data is automatically smoothed and favors straight planes and so forth. Looks very reasonable for automatically captured data. The overlaid video is aligned with the model with image-based post processing of the video, almost entirely alleviating registration errors.

Tuesday

The paper session on VR in medicine was good, but I'll keep it short. The last paper explained a project that combined an HMD with a VR workbench with front projection, so to speak a high-res "magnifying" glass concept with the workbench serving the peripheral view. Surgeons apparently liked this, and it seems to confirm a definite market for this concept. Patrick Baudisch did similar work in his "focus plus context screen" project, only for non-VR desktop systems.

One of the highlights of the conference was the panel discussion about VR vs. AR. Dieter Schmalstieg (TU Wien) was the only panel participant who "quietly" suggested to focus on AR research for now, because VR seems to have stalled a little and could definitely benefit from most AR advances. AR is harder, so there is more room for good research. As one of the steps that could bring us this way he proposed to integrate camera and video support into the VR toolkits ... then the are almost AR toolkits already. Ed Swan (Naval Research Lab) pointed out usability aspects: The best sign of good usability is if users think in the problem domain and not in the interaction domain. How to test for this he didn't say. Ed sees user-centered interaction as the crucial component to future systems, and those lie at the intersection of systems, software, and usability engineering research. In other words, we must collaborate with these disciplines in order to build good systems. Mark Mine (Disney Imagineering) noted that for their applications VR has very welcome properties - mostly better immersion and presence - and thus AR is not only harder but maybe even undesired. Steven Feiner stated that "the real world isn't going to go away - deal with it!". Lastly, the panel members agreed that AR and VR both have to rely heavily on psychology research and on the arts in order to achieve its full potential.

The session on Human Performance in VR was interesting throughout. The papers are in the proceedings, this is only a taster. A controversion study from Ben Lok (UNC) tested performance of handling objects in the real world, virtual world with and without haptic feedback, and with or without visible representation of the own arm and hand. Benjamin Watson (northwestern) investigated object placement in 3D. Results: difficulty and display latency don't like each other. Previewing does not add accuracy. Michael Meehan (UNC) presented results from their SIGGRAPH'02 experiment on the virtual pit. Heart beat and skin conductivity reacted significantly stronger for 50ms display latency than for 90ms latency. Thus, they concluded, the threshold for true immersion is lower than 90ms latency. Shumin Zhai (IBM Almaden) discovered some extensions to Fitts Law for a number of more complex tasks, including locomotion in VRs.

Itsuo Sakane from the IAMAS, Japan, gave the banquet speech. It was good, lots of funky arts projects involving VR and other media, but also rather lengthy.

Wednesday

I skipped the multi-user VR session and attended the tracking session instead: First, a simple but robust machine vision system was used for tracking a pen and a cube based on reflective markers on the objects. Next, Eric Foxlin presented his vision-based pose tracker. A combined hard- and software approach to vision-based tracking of distinct fiducials whose location need not be known a priori, plus inertia sensors, was demonstrated to perform flawless in light of varying lighting conditions, very rapid movement, heavy occlusions and what not. The fiducials are 6inch in diameter, which makes them still distinguishable at 12 feet distance with a 60deg FOV lens (640x480). The paper is great, it's a great read about how to do good research _and_ sell a product that the entire VR community is psyched about. A paper on gesture recognition, made independent from the underlying input device, followed a good idea, but its execution rendered it mostly useless in most people's view.

Haptic Devices and Object Manipulation Session: Aaron Bloomfield presented an interesting classification for maintenance tasks (also mentioned a bunch of taxonomies for hand grasps). 8 taxons for task type: fine motor control, significant arm strength, tactile friction, ..., tool-assisted tasks; and 5 types of forces: force, torque, force+torque, where the force can be aligned and not aligned with the direction of motion or grasp. It's hard to explain. Next, a very sweet video presentation showed a voxel-based collision and friction simulation. Essentially, Koichi Hirota et al. were able to very (very!) realistically simulate (render) how two objects interact. For example, they picked up a virtual cube with a cyberglove-driven virtual hand model. As pressure on the fingers was released, the cube first rotated according to its center of mass, and finally it slid out of the grasp. Best of all, they do all this computation in real time in software. Very neat.

The last session covered applications again. Doug Bowman presented "Virtual SAP", an earthquake simulation program. He compared three user interfaces, one immersive, one desktop, one half-tracked for classroom interaction. Anatole Lecuyer (INRIA) showed HOMERE, a training system for blind people that used spatial sound, haptic feedback through the white cane, thermal feedback, and visual feedback for not entirely blind people. They could explore a real environment first in the lab. A gamepad controller with 3 modes (fwd, stop, bkwd) was used for locomotion. Some good results are in the paper. Lastly, Ed Swan compared a bunch of factors for 3D VR navigation in a battlefield visualization system. Interestingly, he found that exocentric navigation fared better in CAVEs and walls, contradicting previous assumptions in the community. For monoscopic displays, egocentric reference frame seems to be better (in terms of faster target reaching) while for stereo displays exocentric performs better. He cautioned however that some of these results might depend on the display resolution rather than intrinsic parameters.

Concluding Words

I have the proceedings and tutorial script on my desk, just grab it if you want to have a look at it. But please leave it in the lab or otherwise send me email that you took it. Thanks. VR 2003 was a great experience for me. A lot of the big names in VR were (physically) present and very approachable. Some very good papers were presented and the exhibition hall had some new members as well.
Last modified: Tue Apr 01 16:16:33 Pacific Standard Time 2003