A1-JamesAndrews
From CS294-10 Visualization Fa07
[edit] Bad Visualization
Explanation
Taken from Shirley's Fundamentals of Computer Graphics. One of these two images was computer generated; the other is a photograph. The images are intended to show that "for environments that have only matte and mirrored surfaces, the Lambertian/specular assumption works well." To understand this, the viewer is expected to confirm the similarity of the images.
This visualization does not work well for two reasons:
1) Confirming the similarity of these two images can be cumbersome and slow. People are bad at identifying differences in similar, adjacent images, even when those differences are significant -- doing so is challenging enough that people make "spot the differences" game books!
2) The choice of scene makes it hard to evaluate differences detected in the two images. The environment is unfamiliar and abstract, so there is less intuitive understanding of how light should behave in this environment. Without that understanding, it's difficult to evaluate if a deviation from 'correct' is acceptable.
Deconstruction
The data here is the quality of the rendering technique, a quantitative value encoded in the differences between the two images.
Redesign
For my redesign, I have added a new image formed by taking the difference of the two images, marking differences above a threshold as red, and overlaying those markings on the rendered image. This difference map allows me to recognize image differences that would be very difficult to extract otherwise; notably the difference along the right edge of the 'doorway'. It also quickly demonstrates that the regions of substantial difference are small. Finally, using the difference map I can be confident in the qualitative similarity of the images, as I only need to closely inspect the regions which were marked in red.
Unfortunately imperfections introduced in the transfer from pixels to page and back again make the difference map less useful in evaluating the underlying rendering technique, as these errors wash out more perceptually interesting differences in the original images, such as the subtle differences in the light pattern on the left wall & ceiling and the difference in level of black in the doorway. Additionally, a more recognizable scene would have helped viewers evaluate the importance of observed differences, but creating a recognizable scene consisting of only matte and mirror surfaces (and an identical digital copy) could be prohibitively time consuming.
[edit] Good Visualization
Explanation
Taken from David Poole's Linear Algebra: A Modern Introduction. This visualization demonstrates the use of SVD for image compression, presenting an image of Gauss approximated by k singular values for k from 2 to 256. Although it is similar conceptually to the above visualization, it works much better because:
1) It is a familiar scene. Our minds are well-trained to recognize the subtle features of faces, so we are able to judge the naturalness of these images with much greater ease.
2) Similarly, a face is a kind of input we are likely to see in image compression, so the performance of this system for Gauss's face is more clearly relevant.
3) We are shown what bad errors (eg, k=2) look like, and are given a sense of the kind of error that the system can generate. This makes it much easier to quickly judge whether or not a problematic error is likely to be present in the better-looking images.
Deconstruction
This visualization shows quality of the image compression (quantitative data) encoded in the visual difference between the target image (top left) and a compressed image. It shows number of singular values used in creating each image (quantitative data) in a label and also in x&y position on the page.
