Paul Fearing
Predictive Rendering
| Degree: | M.Sc. |
| Type: | thesis |
| Year: | 1996 |
| Supervisor: | Alain Fournier |
| Electronic: |
![[PDF]](/labs/imager/th/icons/pdf.gif){kind=icon} |
| Hardcopy: | 107 pages |
Abstract
Computer graphics has always been concerned with increasing rendering speed. This thesis introduces a new method to reduce the cost of the rendering pipeline. It uses fast, simple transformations to predict scene motion some small number of frames in the future. Scene primitives are grouped into nodes in a "plan tree" according to the predicted future motion of their projected screen coordinates. Different nodes in the tree are rendered at different frame rates. Nodes containing slowly moving or static primitives can be rendered many fewer times than nodes containing quickly moving primitives. The rendered scene subsets are depth composited together to form a final frame.Predictive rendering draws graphics primitives only when they move. This contrasts with usual rendering methods, where all primitives (moving or not) are drawn each and every frame. Depending on the amount of temproal coherence between frames, predictive rendering reduces the number of primitives sent to the graphics pipeline, allowing dramatic improvements in overall rendering speeds.
In scenes with large amounts of temporal coherence, our implementation achieved speedups ranging between 270% and 1,850% of the normal rendering times. Analysis on several test scenes (varying in polygonal complexity) showed worst-case prediction costs that ranged between 0.05% and 13.5% of the total rendering time. Actual implementation showed that worst case costs varied between < 1% to 13%, depending on the scene. These costs indicate the penalty that must be paid in scenes with absolutely no temporal coherence.
@MastersThesis{Fearing1996,
author = {Paul Fearing, M.Sc},
title = {Predictive Rendering},
school = {UBC},
year = {1996},
supervisor = {Alain Fournier},
}