Efficient rendering for three-dimensional displays

Abstract

This thesis explores more efficient methods for visualizing point data sets on three-dimensional (3D) displays. Point data sets are used in many scientific applications, e.g. cosmological simulations. Visualizing these data sets in is desirable because it can more readily reveal structure and unknown phenomena. However, cutting-edge scientific point data sets are very large and producing/rendering even a single image is expensive. Furthermore, current literature suggests that the ideal number of views for 3D (multiview) displays can be in the hundreds, which compounds the costs.

The accepted notion that many views are required for displays is challenged by carrying out a novel human factor trials study. The results suggest that humans are actually surprisingly insensitive to the number of viewpoints with regard to their task performance, when occlusion in the scene is not a dominant factor.

Existing stereoscopic rendering algorithms can have high set-up costs which limits their use and none are tuned for uncorrelated point rendering. This thesis shows that it is possible to improve rendering speeds for a low number of views by perspective reprojection. The novelty in the approach described lies in delaying the reprojection and generation of the viewpoints until the fragment stage of the pipeline and streamlining the rendering pipeline for points only. Theoretical analysis suggests a fragment reprojection scheme will render at least 2.8 times faster than navely re-rendering the scene from multiple viewpoints.

Building upon the fragment reprojection technique, further rendering performance is shown to be possible (at the cost of some rendering accuracy) by restricting the amount of reprojection required according to the stereoscopic resolution of the display. A significant benefit is that the scene depth can be mapped arbitrarily to the perceived depth range of the display at no extra cost than a single region mapping approach. Using an average case-study (rendering from a 500k points for a 9-view High Definition 3D display), theoretical analysis suggests that this new approach is capable of twice the performance gains than simply reprojecting every single fragment, and quantitative measures show the algorithm to be 5 times faster than a naïve rendering approach. Further detailed quantitative results, under varying scenarios, are provided and discussed.