Region impostors, or infinite draw distance

[John-Nagle]

At the request of Geenz Linden, I'm starting this discussion here.

Ref: https://www.animats.com/sharpview/technotes/impostor.html

[John-Nagle]

This is something I have in test in Sharpview, my experimental Rust viewer. It's real enough that you can try it. (Downloading Sharpview requires a password, but if you need access, write me.) This discussion is about using the same technology with other viewers.

The general idea is to combine concepts from the SL map, off-region terrain sculpts, Tyche Shepard's Grid Survey, Open Drone Map, and Google Earth to show the SL world all the way out to the edges of the grid, all the time. It turns out that's not too hard.

Things to know that might concern some people:

• Because there are impostors of multiple regions as one object, only about 80 objects at a time represent the whole SL grid. Some of those objects are very low res representations of as many as 1024 regions. As you move around, which impostors are shown at what scale changes. So this is a low-overhead feature. If you know how a "slippy map" works, it's the same concept.
• Impostors are created from pictures of regions, not the detailed models. They only show land and maybe building exteriors. From an intellectual property perspective, impostor building involves taking pictures, as allowed by the Second Life Terms of Service. Those pictures are then processed. It's a lot like how Google Earth makes low-rez views of the ground.
• Single isolated regions don't need impostors, because you can never view them from the outside. Also, you can only see regions you could walk to. So those big areas of checker boarded private regions don't lose their infinite water views. Private estates don't lose any privacy.
• All imagery and models reside in the SL content delivery network and have UUIDs. The index is currently a file distributed with Sharpview but will move to a web site. Eventually I'd like to see LL take that over.

[Geenz]

Discussing it a little internally - we might be able to provide a bit of support by exposing a field that could store an object UUID that viewers could query. It's unlikely we will be able to dedicate much in terms of hours to this however - but we might be able to at least set the foundation for people to build out the viewer end of this.

Some important details you might want to consider rendering-wise:

• Blinn Phong uses the following material format if you want to inject directly into the deferred renderer:
  • Albedo (RGB) Emissive mask (A)
  • Normal (RGB) glossiness (A)
  • Specular (RGB) environment mask (A)
• PBR if you want to inject into the deferred renderer:
  • Base color (RGB)
  • Normal (RGB)
  • Occlusion, Roughness, Metallic (RGB respectively)
  • Emission (RGB - non-masked unlike albedo in BP, what's in your emission texture is exactly what will be emitted)
  • Cutout alpha for "cut out" transparency (stored in base color alpha - will work in the deferred renderer)

[John-Nagle]

Re layers: Since the impostor info comes from pictures, not the internal SL data, we can only capture some layers. Currently, only the base color and emissive layers are used. The emissive layer is for lights. I'd like to take aerial photos of a region with diffuse daylight, and again with no sunlight at all. Those become the base color and emissive layers for the 3D impostor. The impostors are then lit and rendered with whatever environment is currently active. At night, the impostors would show everything that's lit at night. Buildings. Lighthouses, Pools of light from street lights. Airport runways.

It's mostly to orient users. That's really the point of this. Most of the time, in SL, if you're out in the open and not in a valley, you'll be able to see Mt. Campion and the Snowlands in the distance, giving a sense of place and some points of reference in the world.

The camera can get no closer than the draw distance to an region impostor; get closer and the next region goes live.
There's no texture repeating. All terrain textures are 0..1 UV across the region. With a 512x512 texture, each pixel is 0.5m across. Unless the user turns the draw distance way down, they can't get close to an impostor.

Much of what I'm working on now is to let people see what that looks like. Do we need higher-rez textures? Is 512x512 enough? 1024x1024 might be necessary sometimes.

[John-Nagle]

Re what to store: How to store the info for viewers is an open question. Current plan:

• All textures and meshes are ordinary SL assets. They all come from the content delivery network, not the sim servers. Can't do much with object IDs; they belong to sim servers. Have to work only with texture, sculpt texture, and mesh UUIDs. All impostor objects are as they would appear immediately after upload, with no in-world editing.
• There's an index. Right now, it's a text file of JSON shipped with Sharpview. The link at the top of this page shows the JSON format. Working on a server to distribute the JSON. The whole grid will need maybe 10MB of JSON, uncompressed. At some point, I'd like to have someone else, preferably LL, maintain that.
• There are "visibility groups", indicating which regions are visible from other regions. Those are precomputed. It's a transitive closure on "has an edge in common". All of mainland, now that it's connected, is one visibility group. On login or teleport, but not on region crossing, the visibility group ID of the starting region is fetched. Only impostors in that visibility group will be shown. Single-region estates have no visibility group and get no impostors. That cuts the impostor building workload in half. Large multi-region estates get impostors, although this might be made opt-in.
• For each visibility group, an enclosing rectangle is computed. Within that rectangle, there are all the impostors for that visibility group. The necessary 4, 16, 64, 256, and 1024 region impostors are pre-generated to cover the entire rectangle. Holes, including estates from other visibility groups, are filled with a water area impostor. So users see the same regions they'd see now if they turned the draw distance way up.

Delivering the needed info to the viewer is simple. Making good impostors is hard. The plan is to start with plain terrain impostors, get some viewer support, then gradually upgrade the impostors without further viewer changes. Initial impostors will be just the ground texture atop a terrain sculpt, plus a water level. Should be great for sailors; you'll be able to see the far shore.

I'm tempted to figure out some way to add lighthouses and runways to the impostors as aids to navigation. Other than that, no buildings or trees in this phase.

All of this is down at ground level. Skyboxes will not be impostored. A altitude of 128 meters above the highest terrain point in a region is suggested for picture taking. There are maybe ten towers in SL taller than that.

[John-Nagle]

A useful LL side feature to consider initially is naming but not implementing a capability which tells viewers where to find the impostor index. Without that capability, lookup will go to a server of mine, but LL can preempt that at any time and take over the impostor system.