Three Foveation Technique

We evaluated the impact of latency for three different methods of visual foveation:
    Subsampling. Reducing image resolution with increasing retinal eccentricity, bilinearly upsampled using the texture unit. This method is based on the one proposed by Guenter et al. [13].
    Gaussian Blur. Gaussian image-space blur with blur radius increasing with increasing retinal eccentricity.
    fCPS. The method described by Patney et al. [32], which employs foveated coarse-pixel shading with prefiltered shading, foveated temporal anti-aliasing, and post-processed contrast enhancement.

    All of our experiments used the classroom scene in Figure 2, which shows the same viewpoint without foveation and with foveation applied for each of the above methods (for more information about the spatial frequency characteristics of the scene itself, please see the Appendix). The subsampling technique may produce flicker and jagged edges, and therefore serves as a minimum benchmark for less temporally stable methods. Gaussian blur is a maximally stable upper benchmark, although it does not provide any computational savings and reduces contrast. Finally, the fCPS technique provides an evaluation of recent research in the area of foveated rendering. The visual artifacts induced by the fCPS technique are temporally stable jagged edges, ghosting, and contrast distortion.

    Figure 3 shows our three foveation density mappings where the visual field is divided into three regions based on retinal eccentricity:

    For any given eccentricity, our foveation density mapping returns a blur radius between 0 (foveal) and a maximum (peripheral) chosen differently for each trial. We use the blur radius directly as the standard deviation for the Gaussian Blur method. For the other methods, we compute the subsampling factor as twice the blur radius. Thus, we interpret a blur radius of 1 as a 2 × 2 resolution downscale for subsampling, and a 2 × 2 coarse shading rate for fCPS. Note that following the original proposal, we have limited the maximum blur radius for fCPS to 2, corresponding to a 4 × 4 coarse shading rate.


Reference:
R. Albert, A. Patney, D. Luebke, and J. Kim. Latency requirementsfor foveated rendering in virtual reality. ACM Transactions on Applied Perception (TAP), 14(4):25, 2017.