Point Clouds Are the Technology of the Future (And Always Will Be)

“Unlimited Detail” is back – you can see the videos and read some criticism here.  I have never seen a really good white paper on the technology, so I’m going to have to speculate a bit about what it is they’ve actually done, and then I’ll use the rest of the post to describe why this isn’t the only way to improve perceived realism in a game (and is not the most likely one to succeed.

But first: some video.  This is Euclideon’s promo video, showing lots of really ugly polygonal models, and some clearly not polygonal models with a lot of repeating things.

And here we have in-game footage from the upcoming Battlefield 3, using the Frostbite 2 engine.

I’m starting with the video because I had the same first reaction that I think a lot of other 3-d graphics developers had: attacking six-sided trees from Crysis 1 is a straw man; the industry has moved beyond that.  Look at BF3: is the problem really that they don’t have enough polygons in their budget?  Do you see anything that looks like a mesh?

What Is Unlimited Detail, Anyway?

Short answer is: I don’t know – the company has been quite vague about specific technical claims.  This is what I think is going on from their promotional material.

Their rendering engine uses point cloud data instead of shaded, mapped, textured “polygonal soup” as the input data to be rendered.  Their algorithm does high performance culling and level of detail on the highest level of point cloud data. (Whether this is done by precomputing lower-res point  clouds for farther views, like we do now for textures and meshes, is not specified.)

Why Are Polygons Limiting?

First, we have to observe the obvious: a 1080p video image contains a bit over 2 million pixels; today’s video cards can easily draw 2 million vertices per frame at 30+ fps (even over theAGP bus!).  So for a modern GPU, polygon count is not the operative limit.  If you add more polygons, you can’t see them because they become smaller than one pixel on the screen.

The limit for polygons is level of detail.  If the polygonal mesh of your model is static, then when you walk away from it, the polygons are (in screen space) too small (and we run out of polygons – if we are drawing more than one vertex per screen pixel, we can exceed budget) and if we move in, the polygons are too big.

In other words, the problem with polygons is scalability, not raw count.

And in this, Euclideon may have a nugget of truth in their claims: if there is a general purpose algorithm to take a high-polygon irregular 3-d triangle mesh and produce a lower LOD in real time, I am not aware of it.  In other words, you can’t tell the graphics card “listen, the airplane has a million vertices, but can you just draw 5,000 vertices to make an approximation.”  Polygons don’t work like that.

Coping With Polygon Limit: Old School Edition

There’s a fairly simple solution to the problem of non-scalable polygons: you simply pre-create multiple versions of your mesh with different polygon counts – usually by letting your authoring system do this for you.*  X-Plane has this with our ATTR_LOD system.  It’s simple, and it works sort of.

The biggest problem with this is simple data storage.  I usually advise authors to store only two LODs of their models because each LOD takes up storage, and you get limited benefit from each one.  Had really smooth LOD on objects been a design priority, we could have easily designed a streaming system to push the LODs of an object out to disk (just like we do for orthophoto textures), which would allow for a large number of stored LOD variants.  Still, even with this system you can see that the scalability is so-so.

There’s another category of old-school solutions: CPU-generated dynamic meshes.  More traditional flight simulators often use an algorithm like ROAM to rebuild meshes on the fly at varying levels of detail.  When the goal is to render a height field (e.g. a DEM), ROAM gives you all of the nice properties that Euclideon claims – unlimited detail in the near view scaled out arbitrarily far in the far view.  But it must be pointed out that ROAM is specific to height fields – for general purpose meshes like rocks and airplanes, we only have “substitution LOD”, and it’s not that good.

Don’t Repeat Yourself

It should be noted that if we only had to have one unique type of house in our world, we could create unlimited detail with polygons.  We’d just build the house at 800 levels of detail, all the way from “crude” to “microscopic” and show the right version.  Polygonal renderers do this well.

What stops us is that the mesh budget would be blown on one house; if we need every house to be different, LOD by brute force isn’t going to work.

That’s why the number of repeating structures in the Euclideon demo videos gives developers a queasy feeling.  There are two possibilities:

1. When they built their demo world, they repeated structures over and over because it was a quick way to make something complex, then saved the huge resulting data set to disk.
2. They stored each repeating part only once and are drawing multiple copies.

If it’s the second case, that’s just not impressive, because games can do this now – it’s called “instancing“, and it’s very high performance.  If it’s the first case, well that was just silly – if their engine can draw that much unique detail, they should have filled their world with unique “stuff” to show the engine off.

Where Does Detail Come From?

Before we go on to how modern games create scalable polygonal meshes, we have to ask an important question: where do these details come from?

The claim of infinite detail is that you would build your world in ridiculously high resolution and let the engine handle the down-sampling for scalability automatically.  I don’t think this is a realistic view of the production process.

For X-Plane, the limit on detail is primarily data size.  We already (for X-Plane 9) ship a 78 GB scenery product.  But it’s the structure of that detail that is more interesting.

The scenery is created by “crossing” data sets at two resolutions to create the illusion of something much more detailed.  We take the mesh data (at 90m or worse resolution) and texture it with “landclass” textures – repeating snippets of terrain texture at about 4 meters per pixel.  The terrain is about 78 GB (with 3-d annotations, uncompressed) and the terrain textures are perhaps 250 MB.  If we were to simply ship 4 meter per pixel orthophotos for the world, I think we’d need about 9.3 trillion pixels of texture data.

I mention this because crossing multiple levels of detail is often both part of an authoring process (I will apply the “scales” bump map to the “demon” mesh, then apply a “skin” shader) and how we achieve good data compression.  If the “crossed” art elements never have to be multiplied out, we can store the demon at low res, and the scales bump map over a short distance.  There can be cases where an author simply wants to create one huge art asset, but a lot of the time, large scale really means multiple scale.

Coping With Polygon Limit: New School Edition

If we understand that art assets often are a mash-up of elements running at different scales, we can see how the latest generation of hardware lets us blow past the polygon limit while keeping our data set on disk small.

• DX11 cards come with hardware tessellation.  If our mesh becomes detailed via a control mesh, curve interpolation (e.g. NURBS or whatever) and some kind of displacement mapping, we can simply put the source art elements no the GPU and let the GPU multiply it out on the fly, with variable polygon resolution based on view angle.  Since this is done per frame, we can get the right number of polygons per frame.**
• Since DX10 we’ve had reasonably good flow control in shaders, which allow for displacement mapping and other convincing promotion of 2-d detail to 3-d.

So we can see a choice for game engine developers:

• Switch to point cloud data and a new way of rendering it.  Use the art tools to generate an absolutely ginormous point cloud and trust the scalability of the engine or
• Switch to DX11, push the sources of the art asset to the GPU, and let the GPU do the data generation in real-time.

The advantages of pushing the problem “down to the GPU” (rather than moving to point clouds) is that it lets you ship the smaller set of “generators” for detail, rather than the complete data set.

Euclideon does mention this toward the end of their YouTube video, when they try to categorize art assets into “fiction” (generated by art tools) and “non-fiction” (generated by super-high-resolution scanners).

I don’t deny that if your goal is “non-fiction” – that is, to simply high-res scan a huge world and have it in total detail, not even clever DX11 tricks are going to help you.  But I question how useful that is for anyone in the games industry.  I expect game worlds to be mostly fiction, because they can be.

If I build a game world and I populate my overpasses with concrete support pylons, which am I going to do?

• Scan hundreds of thousands of pylons all around San Diego so I can have “the actual concrete”? or
• Model about 10 pylons and use them over and over, perhaps with a shader that “dirties them up” a little bit differently every time based on a noise source?

There are industries (I’m thinking GIS and medical imaging) where being able to visualize “the real data set” is absolutely critical – and it may be that Euclideon gains traction there.  But for the game development pipeline, I expect fiction, I expect the crossing of multiple levels of detail, and I expect final storage space to be a real factor.

Final Id Thoughts

Two final notes, both regarding Id software…

John Carmack has come down on the side of “large art assets” as superior to “procedural generation” – that is, between an algorithm that expands data and having the artists “just make more” the later is preferable.  The thrust of my argument (huge data sets aren’t shippable, and the generators of detail are being pushed to the GPU) seems like it goes against that, but I agree with Carmack for the scales he is referring to.  Procedural mountains aren’t a substitute for real DEMs.  I think Carmack’s argument is that we can’t cut down the amount of game content from what currently ships without losing quality.  My argument is we can’t scale it up a ton without hitting distribution problems.

Finally, point clouds aren’t the only way to get scalable non-polygonal rendering; a few years ago everyone got very excited about Sparse Voxel Octrees (SVOs).  A SVO is basically a 3-d texture with transparency for empty space, encoded in a very clever and efficient manner for fast rasterization and high compression.  Will SVOs replace polygons?  I don’t know; I suspect that we can make the same arguments against SVOs that we make against point clouds.  I’m waiting to see a game put them into heavy use.

* E.g. the artist would model using NURBS and a displacement map, then let the 3-d tool “polygonalize” the model with different levels of subdivision.  At high subdivision levels, smooth curves are smooth and the displacement map provides smaller detail.

** The polygon limit also comes from CPU-GPU interaction, so when final mesh generation is moved to the GPU we also just get a lot more polygons.

About Ben Supnik

Ben is a software engineer who works on X-Plane; he spends most of his days drinking coffee and swearing at the computer -- sometimes at the same time.

View all posts by Ben Supnik →