Visual Fidelity

Measurement and Prediction

The need for interactivity forces visual quality compromises. We would like to preserve visual fidelity as much as possible as we make those compromises. Unfortunately, we generally don't have time to ask users what they think about the results we produce. We are studying the relationship between perceived visual quality, and various methods of automatically predicting that visual quality. To date, it seems fairly simple approaches work quite well. In the process, we are also performing basic research about perception itself.

Project members: Ben Watson (NU), Alinda Friedman (U Alberta) & Aaron McGaffey.

Sponsors: NSF award 0093172.

Data, Results & Testbed

We are making the 3D stimuli, 2D stimuli and resulting measures of those stimuli from our SIGGRAPH work (referenced below) available for use by the research community as a testbed. All 2D stimuli and measures are provided, but a few of the 3D stimuli were obtained from a commercial source -- we are not able to distribute those. Fortunately, the prices for those models are very reasonable.

3D stimuli -- (ply)
2D stimuli -- (tif)
measures -- (MS Excel)

If you have any questions about these files and their formats, please contact us.

Publications

B.A. Watson, A. Friedman, A. McGaffey (2001). Measuring and predicting visual fidelity (pdf). Talk (ppt.zip, pdf). Proc. SIGGRAPH 2001 (Los Angeles, August). In Computer Graphics Proceedings, Annual Conference Series, ACM SIGGRAPH, 213-220.

B.A. Watson, A. Friedman, A. McGaffey (2000). Using naming time to evaluate quality predictors for model simplification (pdf). Proc. ACM Computer Human Interaction (CHI), 113-120.

Imagery

	Naming times and simplification We can measure perceived visual quality by asking people to name what they see, and recording the time that takes. Here simplification makes things harder to name, and QSlim is better at preserving quality than vertex clustering.
	How effective are automatic fidelity predictors? Even fairly simple predictors are quite good at predicting perceived visual fidelity -- as measured with subjective ratings or choice. Naming time results are harder to predict. Here, red indicates poor automatic/experimental correlations.