Here is a link to the PDF of our paper, Detail to Attention: Exploiting Visual Tasks for Selective Rendering by Kirstin Cater, Greg Ward, and Alan Chalmers.
We have provided a
separate 2-page document containing the color
figures from the paper for the viewer’s convenience.
The images
for the experiment were created using Radiance, and used two models of an
office scene the only difference being the location of items in the scene
mainly teapots.
See HQSceneA.jpg and HQSceneB.jpg.
Both of these images were rendered at a resolution of 3072x3072, which is what
we refer to as High Quality (HQ) in the paper.
The other
two rendering qualities we used in the experiment were Low Quality (LQ) where
the image was rendered at a resolution of 1024x1024 -- see images LQSceneA.jpg and LQSceneB.jpg,
and Selective Quality (SQ) -- see images SQSceneA.jpg
and SQSceneB.jpg.
The selective quality images were created by selectively rendering the majority of the scene in low quality (resolution of 1024x1024) apart from the visual angle of the fovea (2 degrees) centred on each teapot. Also included is an example of an eye scan tracked when an observer was counting the teapots, see image EyeTrackTeapot.jpg, and another eye scan superimposed on the VDP image calculated -- see VDPEyeTrack.jpg.
Here is the first part
of the animation at 320x240 resolution with
the highest quality settings, so the artifacts you see should be due to our
rendering (as opposed to compression).
You may need to copy this file to your hard disk to get an adequate
display if your CD-ROM is not particularly fast.
The AVI file seg1bonus.avi shows our animation in the upper left
quadrant, with corresponding initial error estimate after IBR refinement in the
upper right, the conspicuity map in the lower left, and the number of final
samples in the lower right. We offer this as insight to the behavior of our
method in operation. On occasion, these maps will flash at places where we had
a break in our segment, or had to restart the animation after some problem when
it didn’t have the prior frame to work from.