My research in perception and graphics has been motivated by the goal of using knowledge of perception -- and thus display fidelity -- to improve rendering for interactive computer graphics. In this effort, I found myself addressing many of the questions not answered (at least not directly) in existing perceptual results, and have performed many of my own experiments in collaboration with psychologists. Some of the results are of interest for non-interactive as well as interactive graphics applications. These collaborations have been fairly unique, involving engineering and cognitive psychologists rather than psychophysicists.

I believe that research using the goal-directed and top-down approaches of engineering and cognitive psychologists form a crucial complement to the psychophysically-based studies and systems described to date. As the organizers of this campfire have noted, the applied sorts of questions we graphics researchers tend to pose involve very many experimental variables, making it difficult to generalize from highly controlled psychophysically oriented results. In recognition of this, psychology has already made room within itself for “high-” and “low-level” psychology.

The first series of experiments we performed addressed the effectiveness of reducing detail in the periphery of interactive display, leveraging the drop in visual sensitivity with eccentricity. Participants performed a demanding search task that required heavy use of display peripheries. Results were encouraging [Wats97a]. Perhaps the most interesting result addressed the effectiveness of using only head tracking rather than eye tracking with this display technique. We found that modeling a head tracker as an eye tracker with 15 degrees of error worked quite effectively [Wats97b]. This confirmed previous results from the psychological literature in our applied graphics context.

An additional series of experiments examined the importance of temporal detail, in the form of frame rate and delay. In several studies [Wats97c, Wats98], we examined the relationship of temporal detail to the performance of several tasks. Among many other things, we found that temporal detail is more important in tasks that require constant feedback (e.g. placing something) than in tasks that do not (e.g. shooting). We also found that variation in temporal detail is relatively unimportant, though users are not completely insensitive to it.

These results provide guidelines for detail and fidelity management. However, they do not specifically address the tradeoff of visual and temporal detail. We began to tackle this issue by introducing the concept of I/O differencing. The basic insight is that delay of any sort introduces spatial inaccuracies, just as visual approximation does: objects are displayed where they *were*, rather than where they *are*. We implemented and experimented with a 3D rotation technique using I/O differencing in [Meru99]. We believe that although I/O differencing is a simple insight, it is a powerful one, and our experiments with it continue.

Finally, we have confronted the issue of visual fidelity measurement directly in a collaboration with cognitive psychologists [Wats00]. In our latest experiment, we display 3D models to participants and obtain three experimental measures of quality (naming time, ratings, and forced choice) in response to manipulations of three experimental variables (animal or object model, and type and amount of model simplification). We then used these results to evaluate three automatic measures of quality and fidelity (image MSE, the measure from Bolin and Meyer [Boli98], and 3D distance [Cign98]). The results are surprising and provocative, indicating among other things that at least for these stimuli, MSE is as good as the much more involved Bolin and Meyer measure. Perhaps more important, none of the automatic measures was very successful in predicting naming time ­- the measure which most closely indexes the process of recognition.

Our current research efforts are directed towards implementation of I/O differencing in a more complex 3D context, and following up our experimental results on fidelity measures with experiments using more varied stimuli, and ultimately improved automatic fidelity measures.

We believe the following questions deserve discussion at the campfire:

• How should visual fidelity be experimentally measured?
• For interactive applications, how can temporal fidelity be experimentally measured?
• How can automatic measures of fidelity be improved?
• How can fidelity be automatically measured during real time applications?
• What does fidelity mean in unnatural settings such as information visualization?

References