18 minutes ago, NoAbsoloops said:

 That's very frustrating.

Does this happen often on the internet?

 

Yeah in general tempers do tend to flare on the internet. I think it's exacerbated by anonymity, and lack of face to face contact.

1 hour ago, NoAbsoloops said:

I know what NN's are. That's why I introduced them in the context of sophisticated rendering or physics engine techniques. In fact, I brought up an article about it?

On 6/3/2018 at 3:11 AM, NoAbsoloops said:

This is what lead me to believe that machine learning can really amplify physics and rendering.
It can really scale up physics engines I think.

I literally did. and then you tell me that I say that they're only limited to 3d voxel points? wtf? i don't get that, That's very frustrating.

Why offense?

However, NNs have been used with good results to mimic fluid simulation. Also in control problems like character / locomotion simulation. I don't know of any application for rigid body simulation.

In rendering it has been used to mimic effects like SSAO or DOF, but slower and worse quality. More promising is recent progress in using NN to denoise path traced images with low sample count. But the approaches not using NNs seem at least as good / fast.

You typically use NNs for problems where you don't know how to solve them, or the afford would be too large, like image detection. It's no magic bullet.

For destruction and large worlds, voxels etc. state of the art, you can look at https://www.voxelfarm.com/index.html and https://www.atomontage.com/ but you probably have heard about this already.

I've been trying to get a good understanding of what NN's can generally improve, thanks. 

I have not heard of those actually. Thanks.

This is a bit concerning. 
I hope compression works well.
I think clients will also already "Know" a lot about the voxel data before deltas/changes are applied on the client side.

3 hours ago, NoAbsoloops said:

This is a bit concerning. 

Most people, including myself, do not believe voxels will replace polygon meshes entirely. Polygon meshes are much more optimal:

They cover only the surface, because we are interested only in the surface. If we want destruction, it's more efficient to generate new polygons on the fly, than to have all the volumetric data available just in case it becomes visible due to destruction.

Content creation is much easier with polygons as well, for similar reasons. This is not just because polygon tools are better developed at the moment.

The seperation of geometry and texture we use with poly meshes allows reuse / compress of both of them seperately, which is another form of optimization exclusive to poly meshes.

This is said to turn the arguments they make in the problem section against them. Polygons and textures take less memory and less cycles, no matter what compression you use for them. The main downside of polygons is that they hare more complicated and harder to do than voxels. For example if you want displacement mapping on polygons, you need a seamless global UV parametrization to avoid expensive stitching on seams. This is hard to do, and the reason we do not see much displacement mapping yet (it's not performance as most people think). Those algorithms are currently used in industrial design, but did not made it into games yet.

The only kind of geometry where voxels will beat polygons is high frequency vegetation, like a bush with thousands of fine branches and leaves.

This is why something like Voxelfarm should be much more interesting for you, because it converts volume data (or given meshes as well) into 'destructible' polygon meshes. (They use dual contouring, which is similar to marching cubes but preserves sharp features, e.g. the corner of a wall.)

Voxelfarm has already been used in a large scale MMO, 'Everquest' IIRC, but the game has been canceled.

(I made the Automontage link because it's the most feature rich voxel engine i know about. This stuff is interesting to follow, but i'm sure they'll fail with their claim to replace polyogns, as Unlimited Detail did.)

VoxelFarm looks great. I should be able to get started with it.
I saw the video and was wondering how they did destruction.
Thanks. This is more than enough of a starting point for me.

23 minutes ago, NoAbsoloops said:

I saw the video and was wondering how they did destruction.

You can think of it as slicing all geometry by a 3D grid, so you get blocks similar to Minecraft but with a more detailed shape so the grid is not obviously visible. This allows for destruction and also construction in real time.

Here is the game i've meant: https://www.youtube.com/watch?v=MrVMhtrfZVU

I like this one. It feels like Rambo!

that's gorgeous.

Do all the little pieces that break off stay in sync with the server and other client experiences for things like online games?

13 hours ago, JoeJ said:

This is why something like Voxelfarm should be much more interesting for you, because it converts volume data (or given meshes as well) into 'destructible' polygon meshes. (They use dual contouring, which is similar to marching cubes but preserves sharp features, e.g. the corner of a wall.) 

The way I understand dual contouring is that it can't really preserve sharp corers all the time. If geometry is manhattan and aligned with the grid it can for sure, however there can be a lot of cases where the required point would have to lie outside the box so to keep it inside you have to distort your geometry again. You can imagine a corner at 45 to the grid that punches though one side of the box.

I've been trying to think of some algorithm that can preserve sharp corners correctly all time time.  I think it would require that an edge allow two crossings instead of one, and of course you can't just use interpolation to find the crossing points.  I haven't really come up with anything yet though.

I also noticed Everquest Next was canceled after a lot of hype and going though a bunch of contortions, so I'm wondering if there is some underlying problem with the technology. Perhaps the performance wasn't up to snuff.  Do you know if there are any games that have been released that actually use Voxel Farm?

5 minutes ago, Gnollrunner said:

I've been trying to think of some algorithm that can preserve sharp corners correctly all time time.  I think it would require that an edge allow two crossings instead of one, and of course you can't just use interpolation to find the crossing points.  I haven't really come up with anything yet though.

I remember quite a number of papers going beyond dual contouring or extending it. However at this complexity looking at hexagonal meshing might be more interesting. (Although the grid you utilize for coll. det. gets lost there.)

10 minutes ago, Gnollrunner said:

I also noticed Everquest Next was canceled after a lot of hype and going though a bunch of contortions, so I'm wondering if there is some underlying problem with the technology. Perhaps the performance wasn't up to snuff.  Do you know if there are any games that have been released that actually use Voxel Farm?

No, i do not know any game. It seems most teams implement such tech on their own if needed. But i wonder about it, Voxel Farm seems very solid. I guess we'll see some games now with UE4 integration...

I don't think Everquest had performance issues, there are many videos and they look fine. But construction using small fragments as seen in my link is too tedious for users. (One would need to offer larger modules of buildings, create walls from a quick outline etc. to make it easier... some work!)