Demoing stereoscopic image pairs stacked one on top of the other is a peculiar choice. How exactly is one supposed to evaluate that? Cross their eyes up and down?
Top down is the usual way to format 3D equirectangular photos. You need a 360 photo viewer to view them properly.
A thought: if you take as input a video or stills instead of an already stitched panorama, you could take advantage of the parallax between frames to get 3D info. Especially if the user deliberately captures the images on a large arc (hold the camera at arms length and swing around rather than pivoting the camera in place).
This build has been quickly put together (UI, etc) after working mostly on the core functionality for a long time. I've done quite some experimenting too with support for videos, but haven't gotten it to work yet (productionwise) due to constraints of CPU and GPU (Cuda). You'd need a beefy PC but it's possible.
Turning the output into a point cloud and changing the point of view is a better way to display the results. Often, methods that look good with the 2D depth-heatmap give pretty bad point clouds.
I see it only works on Windows :( Are you planning on compiling it for other systems, or even as a webapp?
By the way, does the computer running it requires CUDA?