Quoted from "Klaus"
what are layered depth maps?
A layered depth map represents the appearance of a 3d space from a range of viewpoints, rather than just one; but a limited range, for example a sphere 1m in diameter. It is a specialized 3d model good for rendering high resolution stereoscopic views with parallax shift and occlusion, as might be seen by a stationary observer turning the head and tilting the body. Perfect for VR stills and video frames. In the context of CG rendering, the "height map" is a similar idea.
This representation consists of several overlapping partial models, each with an associated texture. Each part covers a restricted range of depths. In VR the layers will be spherical shells. The textures have alpha channels, so a view from any point in the feasible range can be rendered simply by shifting the models and alpha blending the projected textures. In principle, a simple generalization of single-depthmap, single-texture process you have already implemented.
This idea was developed 20 years ago by the Szeliski group (https://www.colinzheng.com/wp-content/dat…/ldp_cvpr07.pdf ) but not much used outside CG (http://frederikaalund.com/wp-content/upl…-Depth-Maps.pdf ) until recent progress made it feasible to create layered representations from photos. Now this is quickly becoming the preferred format for displaying 'lightfield' and '6-dof'images (https://augmentedperception.github.io/deepviewvideo/ ) (http://visual.cs.brown.edu/projects/matryodshkawebpage/ ). Several other groups have published work targeting layered depth maps. At the moment each group has a different way of packaging layered views, but clearly this can be done simply by combining existing formats and software.
It turns out that 5 or 6 layers is enough, because layers mainly encode visibility, the actual depths being attached to the model vertices. Where there is continuous depth, alpha blending gives good visual continuity because in those places the depth difference between adjacent layers is small. The total data size can be considerably less than 5 or 6 times that of a single layer representation because only the occluded areas need to be multiple. I can imagine that smart tiling and indexing might yield LDM files only about twice the size of a single-layer pano with depth map.
I look forward to being able to package 6-dof VR experiences with krpano.
Kind regards,
Thomas