How far away are we from this goal?
svendr...@hotmail.com
It was said in one of my old text books that in order to achieve compelling
realism with 3D modelling of some day-to-day envoironment, a minimum of 80
million polygons would be needed. (Don't ask me how the author drew this
conclusion.)
I just wonder given the pace of hardware evolution, how long would it take for
this kind of modelling to be feasible on a PC, interactively (aka real-time)?
-----== Posted via Deja News, The Leader in Internet Discussion ==-----
http://www.dejanews.com/ Now offering spam-free web-based newsreading
Mike Anderson
Dunno where 80 million comes from..... but as I see it the problem isn't the
number of polygons but what you can draw on them. Realistic 3D modelling of
general scenes requires proper shadow calculations, semi-transparent objects
and reflections on arbitrary surfaces. It's the calculations for these
rather than the actual number of polygons that are way beyond current
computing power.
I'd say about 6 years until we have photorealistic games and 15-20 years
before we can render arbitrary scenes (raytraced quality) in realtime. Any
other offers?
Mike.
svendr...@hotmail.com wrote in message
<6mdl4c$qhq$1...@nnrp1.dejanews.com>...
Rolf Magnus
svendr...@hotmail.com wrote:
>
> It was said in one of my old text books that in order to achieve
compelling
> realism with 3D modelling of some day-to-day envoironment, a minimum
of 80
> million polygons would be needed. (Don't ask me how the author drew
this
> conclusion.)
>
> I just wonder given the pace of hardware evolution, how long would it
take for
> this kind of modelling to be feasible on a PC, interactively (aka
real-time)?
The voodoo 2, that is afaik the fastest accelerator fof the average home
pc,
is able to draw about 90 million *pixels* per second. How many polygons
it can draw ? I don't know, but I think about 1 million. But you have to
consider that the 3d realtime hardware is advancing with an extremely
speed,
so it may be possible in just a few years.
Rolf Magnus
please remove X to respond via email
(sorry for that, but I was getting more
spam mail than real mail)
krom...@my-dejanews.com
In article <6mdoic$9m7$1...@pegasus.csx.cam.ac.uk>,
"Mike Anderson" <mik...@dial.pipex.com> wrote:
>
> Dunno where 80 million comes from..... but as I see it the problem isn't the
> number of polygons but what you can draw on them. Realistic 3D modelling of
> general scenes requires proper shadow calculations, semi-transparent objects
> and reflections on arbitrary surfaces. It's the calculations for these
> rather than the actual number of polygons that are way beyond current
> computing power.
>
> I'd say about 6 years until we have photorealistic games and 15-20 years
> before we can render arbitrary scenes (raytraced quality) in realtime. Any
> other offers?
>
I'd say 4-5 years before we can render all the polygons we need with
as many effects and textures as we want... And another 5 years after that
before there is a good base of people with talent to use all this power.
I'm not sure how you distinguish raytraced-quality from photorealistic,
but I'm talking Toy-Story type graphics in 5 years.
Andy K.
> Mike.
>
> svendr...@hotmail.com wrote in message
> <6mdl4c$qhq$1...@nnrp1.dejanews.com>...
> >It was said in one of my old text books that in order to achieve compelling
> >realism with 3D modelling of some day-to-day envoironment, a minimum of 80
> >million polygons would be needed. (Don't ask me how the author drew this
> >conclusion.)
> >
> >I just wonder given the pace of hardware evolution, how long would it take
> for
> >this kind of modelling to be feasible on a PC, interactively (aka
> real-time)?
> >
Tagore Smith
Say that you are rendering at 1024x768. That is 786432 pixels. So
from a given viewpoint in the scene the most polygons that are likely to
be useful is under one million. Even using sub-pixel sampling with, say,
twenty-five samples per pixel, you would need less than twenty-million
polygons. Of course as the viewpoint changes different sets of one
million polygons are useful. There currently exist very interesting
algorithms for adaptive mesh decimation. Do a search for Hugues Hoppe,
who has a homepage at Microsoft research.
I work on an Onyx InfiniteReality, which is a fairly powerful real
time renderer. On this machine, using Iris Performer, which does
implement adaptive level of detail, the obstacles to photo-realistic
rendering are not really number of polygons. As Mike Andersen points out
gouraud shading and phong lighting are not terribly photo-realistic. To
get really photo-realistic renderings requires a hybrid raytracing,
radiosity method, which is extremely expensive. Even in the context of
gouraud shading there are real problems with things like texture
distortion when using mip-mapped textures. I would imagine that
inexpensive hardware that will render as many polygons as are useful
will be available within five years.
Tagore Smith
Link
>realism with 3D modelling of some day-to-day envoironment, a minimum of 80
>million polygons would be needed. (Don't ask me how the author drew this
>conclusion.)
>
>I just wonder given the pace of hardware evolution, how long would it take
for
>this kind of modelling to be feasible on a PC, interactively (aka
real-time)?
Not being properly scientific here, but I think its possible now in some
ways. It depends on what you display in your scene. I think it would be
possible to create a 3D scene that looked photo-real if you're careful about
exactly what kinds of objects you're rendering. For example, a realistic
scene of a plain desk-lamp and a beach ball might be possible in real-time
now -- provided the camera doesn't get too close, and the viewer isn't
allowed to look around the room at other stuff :) But doing a simulation of
Brando dribbling water on his bald forehead in a dark, humid jungle... maybe
never.
Even in film, where rendering capabilities are ahead of videogame rendering
by, what?... 15 years?.. still requires a lot of photographic source images
to make up much of the scene. Even with those SGI reality-boxes, it just
gets too expensive to model all the details you see in real life.
- Mike
Brandon Van Every
svendr...@hotmail.com wrote in message
<6mdl4c$qhq$1...@nnrp1.dejanews.com>...
>It was said in one of my old text books that in order to achieve compelling
>realism with 3D modelling of some day-to-day envoironment, a minimum of 80
>million polygons would be needed. (Don't ask me how the author drew this
>conclusion.)
>
>I just wonder given the pace of hardware evolution, how long would it take
for
>this kind of modelling to be feasible on a PC, interactively (aka
real-time)?
Well, how fast is the 3D speed really improving? Some say we're doubling
every 6 months, but I really don't believe that. Main memory is still a
huge bottleneck and it's only gone up a factor of 1.5 in 5 years. And
people keep using the same old shitty source code for their drivers, ducking
lotsa needed long-term improvements because they're on a 6 month product
cycle, so that slows things down. Maybe it's fair to say we're doubling
every 12 to 18 months? At least I'd take those as the minimum/maximum
bounds.
Mainstream commodity hardware does 0.5 million triangles per second today
(in the real world, not in marketroid speak). I'm not talking the best of
the bleeding edge, I'm talking commonly deployed hardware. log2(80/0.5) <
log2(256) = 8 doublings. So 8 to 12 years, thereabouts.
Cheers,
Brandon Van Every
CACDBA
>It was said in one of my old text books that in order to achieve compelling
>realism with 3D modelling of some day-to-day envoironment, a minimum of 80
>million polygons would be needed. (Don't ask me how the author drew this
>conclusion.)
A reaction from a maybe slightly different angle: I'd say that 80
million polygons is far from enough to achieve a convincing realism.
Being concerned with hyper-realistic imaging of exactly these
day-to-day environments I still am amazed every day how much the
'reality' of the image I make is determined by objects that are
outside that scene. F.e.: How much polygons are in the buildings and
trees that are reflected in the chrome of a car-bumper?
CACDBA
*c my work at http://www.xs4all.nl/~sas2*
Jeff Petersen
> A reaction from a maybe slightly different angle: I'd say that 80
> million polygons is far from enough to achieve a convincing realism.
> Being concerned with hyper-realistic imaging of exactly these
> day-to-day environments I still am amazed every day how much the
> 'reality' of the image I make is determined by objects that are
> outside that scene. F.e.: How much polygons are in the buildings and
> trees that are reflected in the chrome of a car-bumper?
Those reflections are not done with seperate polygons per se, but with
reflections, which would require ray-tracing. In order to achieve true
realism, I think you need to move away from polygon based modeling entirely
and have the scene rendered directly from calculations on the curved
surfaces that really make up objects. I don't believe even the fastest
computers in the world can do ray-tracing of complex (lots of curves and
objects and lights) scenes in real-time..so I say we have a very long way
to go.
Defee Pawel
In article <01bd9bd8$6ea6ac60$87438cd1@jeffwork>,
Jeff Petersen <je...@XXXinconnect.com> wrote:
>Those reflections are not done with seperate polygons per se, but with
>reflections, which would require ray-tracing. In order to achieve true
>realism, I think you need to move away from polygon based modeling entirely
>and have the scene rendered directly from calculations on the curved
>surfaces that really make up objects. I don't believe even the fastest
>computers in the world can do ray-tracing of complex (lots of curves and
>objects and lights) scenes in real-time..so I say we have a very long way
>to go.
Just to put my 2 cents, I think you guys are overly concentrating on
classical ray-trace effects like silver ball reflections etc. Talking
about (photo)realism I guess that these kind of effects play a really
minor role in there. Considering a dnyamic scene from real world, you can
render all the chrome objects and your silver spoons using environment
mapping or something similar. They key issue is the global illumination...
proper, real, realistic shading for isntance adds far more to realism than
a caustic from your glass on the table. You won't notice a missing
caustic, but you will notice a missing shadow. I claim that we will be
able to see photorealistic looking real-time rendering very soon, though
it might miss the (non-noticeable) details. The work done on
importance-based rendering with radiosity/radiance is a very significant
step in my opinion, this is the way to go.
--
*******************************************************************************
Pawel Defee
pawel...@cc.tut.fi
*******************************************************************************
-AW
> >realism with 3D modelling of some day-to-day envoironment, a minimum of 80
> >million polygons would be needed. (Don't ask me how the author drew this
> >conclusion.)
> >
> for
> >this kind of modelling to be feasible on a PC, interactively (aka
> real-time)?
>
> ways. It depends on what you display in your scene. I think it would be
> possible to create a 3D scene that looked photo-real if you're careful about
> exactly what kinds of objects you're rendering. For example, a realistic
> scene of a plain desk-lamp and a beach ball might be possible in real-time
> now -- provided the camera doesn't get too close, and the viewer isn't
> allowed to look around the room at other stuff :) But doing a simulation of
> Brando dribbling water on his bald forehead in a dark, humid jungle... maybe
> never.
>
> Even in film, where rendering capabilities are ahead of videogame rendering
> by, what?... 15 years?.. still requires a lot of photographic source images
> to make up much of the scene. Even with those SGI reality-boxes, it just
> gets too expensive to model all the details you see in real life.
>
> - Mike
I don't suppose you've taken a look at Riven lately?
http://www.cyan.com
http://www1.riven.com
http://www.geocities.com/TimesSquare/Alley/riven.htm
Cyan's done quite a job rendering photorealistic environments... of
course,
they render still images using raytracing, not via a realtime rendering
engine.
As for:
> Even in film, where rendering capabilities are ahead of videogame rendering
> by, what?... 15 years?.. still requires a lot of photographic source images
The only reason the film industry is ahead of the game industry is that
they
only have a need to render still images and then animate them. The
amount of
time required to render those images is of no importance to them, so
they are
able to create much more realistic images (via raytracing as in Riven).
In the game industry, realtime rendering is much more important, so
scenes
are generally much more elementary, and much less convincing.
As for the amount of time it'll take for processing power to be such
that
realtime raytracing of scenes can be accomplished, one of the Cyan
employees
estimated that we will have to wait (if my memory serves me correctly)
anywhere
from 5 to 10 years, at least.
-AW
CACDBA
On 20 Jun 1998 00:13:36 GMT, pa...@cc.tut.fi (Defee Pawel) wrote:
<snip>
>classical ray-trace effects like silver ball reflections etc. Talking
>about (photo)realism I guess that these kind of effects play a really
>minor role in there. Considering a dnyamic scene from real world, you can
>render all the chrome objects and your silver spoons using environment
>mapping or something similar.
Although I agree that shading is more elementary than
reflections/refractions I don't think you can say they play a minor
role here. Not when you do use the word photorealism in the same
sentence:
It's exactly those socalled "non-noticable" things that makes
immediately clear if you are dealing with a computer generated image
or f.e. a photo . Don't get me wrong here: I dont think you should
perse aim for that goal eather: I admire computer-animation exactly
for its realism, but not exactly "photo-realism". I'd call it
"constructivist-realism" or something like that. A convincing model,
"real" from a rationalist point of view.
Only when it obviously tries to look fotorealistic I find it often
poor and a bit childish. It's in this sense that I make my remark that
80 million polygons isn't that much. My point here is that if you try
to make it look like reality itself you will have to build not only
your scene but a whole lot more than that, also things that are behind
the viewpoint and so on.
If you don't ,and that is if you use some "effects" instead, it will
never be that photo-realistic and will always look a bit poor....
Tom Plunket
"Link" wrote:
> Even in film, where rendering capabilities are ahead of
> videogame rendering by, what?... 15 years?..
What?!? Tell me that you think film is rendered in
real-time? Just for reference, they aren't, it's on "film"
which is pre-generated. I've never heard any info on it,
but how long did it take to render all of the CGI in Titanic
or Godzilla? A hell of a lot longer than the 2 hours that
the movie takes... Unless they've got it massively parallel
rendering, with one machine per frame. :)
Movies and video games are doing very similar things, it's
just that video games are doing a lot less of it due to the
need for 30+ fps. If it doesn't matter that it takes 2
hours to render a frame, then you can obviously make it look
better.
tom!
Tom Plunket plu...@eidetic.com
video game/3D Studio geek
Eidetic, Inc. +1 541-383-2410 phone
http://www.eidetic.com/ +1 541-383-8905 fax
C.P.I.
svendr...@hotmail.com wrote:
> realism with 3D modelling of some day-to-day envoironment, a minimum of 80
> million polygons would be needed. (Don't ask me how the author drew this
> conclusion.)
>
> I just wonder given the pace of hardware evolution, how long would it take for
> this kind of modelling to be feasible on a PC, interactively (aka real-time)?
Hmmm... I don't know about this 80 million thing... 2 million can
produce VERY good results if the nature of the rendering schemes is good
(i.e. HIGH Quality Shading, Shadows, Fog, Texture mapping, etc.)
Perhaps he means in a whole scene... Also, if you depend on polys
alone, then you would need a lot of polys -- if you render on, say a
bezier patch, you don't need as many divisions <--- (I believe the
Millenium II renders bezier patches or parametric patches, but that is
just an assumption based on the card's intentions -- still, it IS slow)
I think economics will play a big part in this, though. The technology
itself may be within reach in about 6-8 years. But the only problem
is... Once you've got absolutely realistic rendering in realtime --
where do you go from there? What new things can you market to the
public if you've already reached the pinnacle? The hardware
manufacturers will definitely hold off indefinitely because of this...
Unless some new technology for video cards hits... Then, they'll have
something new to work off of.
- C.P.I. / ASMiNC
Defee Pawel
In article <358C2B...@wwa.com>, C.P.I. <c...@wwa.com> wrote:
>svendr...@hotmail.com wrote:
> I think economics will play a big part in this, though. The technology
>itself may be within reach in about 6-8 years. But the only problem
>is... Once you've got absolutely realistic rendering in realtime --
>where do you go from there? What new things can you market to the
>public if you've already reached the pinnacle? The hardware
>manufacturers will definitely hold off indefinitely because of this...
>Unless some new technology for video cards hits... Then, they'll have
>something new to work off of.
The problem in my opinion is that you will never get "realistic enough".
You can always make the problem harder by the order of magnitudes by
considering advanced light behaviour such as non-RGB colour representation
(ie sampled spectrum), anisotropic (true BRDF) reflection and whatever
more.. quantum effects and such. These are part of our dail life... try
simulating a monitor in a ray tracer :) (electrons hitting the screen and
producing the light).
-pd
Brandon Van Every
HPawn wrote in message <358c85e2...@news.freewwweb.com>...
>I feel if hardware makers like 3dfx would have made the hardware be
>able to do things like phong shade or do simple raytracing with the
>API then we would be seeing even more realistic games.
There is no such thing as "simple raytracing through the API." Raytracing
and radiosity are global illumination problems, the speed of which depend on
both the data and the quality of the algorithms. In contrast, hardware
solutions like 3Dfx solve a very specific part of a local illumination
pipeline.
Cheers,
Brandon Van Every
HPawn
I feel if hardware makers like 3dfx would have made the hardware be
able to do things like phong shade or do simple raytracing with the
API then we would be seeing even more realistic games.
On Fri, 19 Jun 1998 12:21:32 GMT, svendr...@hotmail.com wrote:
>It was said in one of my old text books that in order to achieve compelling
>realism with 3D modelling of some day-to-day envoironment, a minimum of 80
>million polygons would be needed. (Don't ask me how the author drew this
>conclusion.)
>
>I just wonder given the pace of hardware evolution, how long would it take for
>this kind of modelling to be feasible on a PC, interactively (aka real-time)?
>
hugo elias
svendr...@hotmail.com wrote:
> It was said in one of my old text books that in order to achieve compelling
> realism with 3D modelling of some day-to-day envoironment, a minimum of 80
> million polygons would be needed. (Don't ask me how the author drew this
> conclusion.)
Do not confuse ray tracing with photorealism. They are not the same
thing.
It is not just about having mathematically accurate models for lighting
etc.
I have seen roughly drawn paintings that look more photorealistic than
some
of the best raytraced images. Photographs have a certain something that
is
still lacking in CGI.
Take a look at Terragen. The author has thought very hard about what it
takes
to make a scene look photorealistic. Some of the images produced are
certainally
begining to move in a photorealistic direction. Note, though, that it
uses flat
shaded oplygons. It is still very unoptimised, and could be made to run
in real
time on todays hardware.
The amount of photorealism you can get depends on the amount of freedom
you're
willing to sacrifice. It's possible to get a totally photorealistic
scene on a
PC today. Take a photo, texture it onto a polygon, render the polygon to
the screen.
Likewise, I'm sure you could create quite a convincing photorealistic
scene, by
photographing everything in a real scene, and texturing it onto a 3D
model of
the scene. It wouldn't handle reflections of course.
Terragen can be seen at:
http://planetside.base.org
-hugo-
-AW
Take a look at these, then tell me what you think.
http://www.geocities.com/SiliconValley/Bay/1200/lansy2_b.jpg
http://www.geocities.com/TimesSquare/Alley/8867/riven3.jpg
http://www.geocities.com/TimesSquare/Alley/8867/watch.jpg
http://www.geocities.com/TimesSquare/Alley/8867/staingla.jpg
http://www.geocities.com/TimesSquare/Alley/8867/solved3.jpg
http://www.geocities.com/TimesSquare/Alley/8867/riven9.jpg
-AW
Brandon Van Every
John Ryland wrote in message ...
>On Sat, 20 Jun 1998, Brandon Van Every wrote:
>>
>> There is no such thing as "simple raytracing through the API."
Raytracing
>> and radiosity are global illumination problems, the speed of which depend
on
>> both the data and the quality of the algorithms. In contrast, hardware
>> solutions like 3Dfx solve a very specific part of a local illumination
>> pipeline.
>
>That's not to say you couldn't implement a ray-tracer in hardware though.
True, in fact there is at least one ray-tracing chip that I'm aware of now.
But that's not the kind of part you find in commodity 3D graphics cards, or
even most high-end 3D graphics cards. 3D graphics acceleration has evolved
to speed up the local illumination parts of the graphics rendering problem,
not the global.
Cheers,
Brandon Van Every
George van den Driessche
In comp.graphics.algorithms C.P.I. <c...@wwa.com> wrote:
> is... Once you've got absolutely realistic rendering in realtime --
> where do you go from there? What new things can you market to the
> public if you've already reached the pinnacle?
Absolutely surrealistic rendering? More detail than reality? :-)
--
George
John Ryland
> What?!? Tell me that you think film is rendered in
> real-time? Just for reference, they aren't, it's on "film"
> which is pre-generated. I've never heard any info on it,
> but how long did it take to render all of the CGI in Titanic
> or Godzilla? A hell of a lot longer than the 2 hours that
> the movie takes...
This is true. I read that the average frame of toy-story took as long as
the movie itself to render. Have a look at:
http://www.usec.sun.com/sunworldonline/swol-11-1995/swol-11-pixar.html
> Unless they've got it massively parallel
> rendering, with one machine per frame. :)
> Movies and video games are doing very similar things, it's
> just that video games are doing a lot less of it due to the
> need for 30+ fps. If it doesn't matter that it takes 2
> hours to render a frame, then you can obviously make it look
> better.
Ok that's true to an extent, but I think the constraint in video games
isn't the 30fps, but the fact they are interactive (ie real-time systems)
that display according to the input. Movies are like a key-framed
animation where if you used computers in parallel you could be rendering
the last frame well before its ready to be displayed.
John
John Ryland
On Sun, 21 Jun 1998, -AW wrote:
> Take a look at these, then tell me what you think.
>
> http://www.geocities.com/SiliconValley/Bay/1200/lansy2_b.jpg
> http://www.geocities.com/TimesSquare/Alley/8867/riven3.jpg
> http://www.geocities.com/TimesSquare/Alley/8867/watch.jpg
> http://www.geocities.com/TimesSquare/Alley/8867/staingla.jpg
> http://www.geocities.com/TimesSquare/Alley/8867/solved3.jpg
> http://www.geocities.com/TimesSquare/Alley/8867/riven9.jpg
They are very nice. Wow! :)
Shadows help *heaps*
How commonly supported are shadows on 3D cards for the PC?
If done using shadow-buffers, they can be implemented in a hardware
inner-loop, can't they?
John
John Ryland
Peter Ashford
On Sat, 20 Jun 1998 19:56:38 GMT, plu...@eidetic.com (Tom Plunket)
wrote:
>"Link" wrote:
>
>> Even in film, where rendering capabilities are ahead of
>> videogame rendering by, what?... 15 years?..
>
>What?!? Tell me that you think film is rendered in
>real-time? Just for reference, they aren't, it's on "film"
>which is pre-generated. I've never heard any info on it,
>but how long did it take to render all of the CGI in Titanic
>or Godzilla? A hell of a lot longer than the 2 hours that
>the movie takes... Unless they've got it massively parallel
>rendering, with one machine per frame. :)
I checked out info on how starship troopers was rendered, and they
were getting frames taking over 30 minutes each.
I didn't catch the original post, but if they're trying to suggest
that 3D techniques in films are more sophisticated than 3D in games, I
don't really agree - games programmers have to do everything in real
time with less resources than are available to film makers.
Peter.
Defee Pawel
In article <Pine.OSF.3.92.980622...@pigeon.qut.edu.au>,
John Ryland <j.ry...@qut.edu.au> wrote:
>They are very nice. Wow! :)
>
>Shadows help *heaps*
>
>How commonly supported are shadows on 3D cards for the PC?
On PC- not really supported and will not be still for some time, before we have an API
that enables the downloading of full scene graph, objects and stuff to the card.
>
>If done using shadow-buffers, they can be implemented in a hardware
>inner-loop, can't they?
>
Take a look closer at those pictures, those shadows are real "soft" shadows. Shadow
buffer's "hard" shadows dont have much to do with real life.
Besides, the complexity increases drastically with the number of light sources :(.
Michael Chock
hugo elias wrote in message <358D35FC...@virgin.net>...
[snip]
>The amount of photorealism you can get depends on the amount of freedom
>you're
>willing to sacrifice. It's possible to get a totally photorealistic
>scene on a
>PC today.
[snip]
Let me tell you about my solution to the global illumination problem for the
special case where there are no emitters in the scene. It runs at 120fps,
rendering 4.6 trillion primitives per frame...
:-P
- Mike
Link
Tom Plunket wrote in message <3590143a...@news.teleport.com>...
>"Link" wrote:
>
>> Even in film, where rendering capabilities are ahead of
>> videogame rendering by, what?... 15 years?..
>
>What?!? Tell me that you think film is rendered in
>real-time? Just for reference, they aren't, it's on "film"
>which is pre-generated. I've never heard any info on it,
>but how long did it take to render all of the CGI in Titanic
>or Godzilla? A hell of a lot longer than the 2 hours that
>the movie takes... Unless they've got it massively parallel
>rendering, with one machine per frame. :)
Hehe, that's pretty funny... you thought I was under the impression the film
industry uses technology 15 years ahead of videogames. No, I meant that
because film is allowed the luxury of rendering "as long as it takes/as much
as it costs" as opposed to rendering 30fps on the base PC requirements, you
get effects that are ~15 years ahead of videogames. Some of the effects from
Tron and Last Starfighter are practical now for videogames. But not
Titanic. (And considering that Titanic used a lot of models, sets and filmed
backgrounds for the effects, that should tell you something about the
likelyhood of photoreal rendering in future videogames...)
- Mike
Chuck McManis
C.P.I. wrote:
> I think economics will play a big part in this, though. The technology
> itself may be within reach in about 6-8 years. But the only problem
> where do you go from there? What new things can you market to the
> manufacturers will definitely hold off indefinitely because of this...
> Unless some new technology for video cards hits... Then, they'll have
> something new to work off of.
The funny thing about this somewhat naive opinion is that it is
often actually held by hardware manufacturers just before they
are knocked off. Consider the most recent example SGI. SGI could
have made better cheaper faster hardware but it would "interfere"
with the economics of the O2 and Octane machines. Guess what, the
market put together machines based on cheap Intel hardware and
new 3D iron and took a big chunk of the pie right out from under
their noses. Where did that leave SGI? Dead.
So full photorealistic hardware will be available as soon as it
is possible to make it. The existing technology companies may try
to stall it but they will be unsuccessful if the market is open.
Once you have photo-realistic rendering in real time (PRRiRT :-)
you'll start doing head tracking or something like it to boost
apparent resolution. Then there will be shared state, etc.
John Ryland
On 22 Jun 1998, Defee Pawel wrote:
> >How commonly supported are shadows on 3D cards for the PC?
>
> On PC- not really supported and will not be still for some time, before we have an API
> that enables the downloading of full scene graph, objects and stuff to the card.
>
> >If done using shadow-buffers, they can be implemented in a hardware
> >inner-loop, can't they?
>
> Take a look closer at those pictures, those shadows are real "soft" shadows. Shadow
> buffer's "hard" shadows dont have much to do with real life.
With shadow-buffers I'm sure you can make soft shadows. I think I read
about it in "Advanced Animation and Rendering Techniques" by Watt and
Watt.
> Besides, the complexity increases drastically with the number of light sources :(.
That would matter as much if it was done in hardware.
Just thinking about this it seems obvious there is so much in an inner
loop that can be done in parallel. If you were doing texture mapped, phong
shaded polygons with shadows, in parallel you could be working out the
perspective correct texture co-ordinate and looking it up, whilst at the
same time you could be calculating the phong lighting for all the lights,
whilst also simialtaneously calculating the amount the pixel is in shadow
from all of the lights (ie for each light in parallel).
Then you could be blending/adding these values up and displaying the pixel
while you are doing these calculations again for the next pixel.
(of course you could also be doing several pixels at once too, just
depends how much silicon you have to play with)
if n is the number of lights that's 2+2*n things in parallel
John
-AW
John Ryland wrote:
<snip>
> if n is the number of lights that's 2+2*n things in parallel
>
> John
Just as a side note: those pictures I posted links to used anywhere
from 4 to 10 light sources, according to Cyan, Inc.
-AW
Jorrit Tyberghein
Michael Chock wrote in message <6mm19l$fun$1...@client2.news.psi.net>...
>
>hugo elias wrote in message <358D35FC...@virgin.net>...
>[snip]
>>The amount of photorealism you can get depends on the amount of freedom
>>you're
>>willing to sacrifice. It's possible to get a totally photorealistic
>>scene on a
>>PC today.
Is this true? I believe that there are still some areas which are difficult
to render photorealisticly (I believe it had to do with water, snow and such).
I thought it was a matter of not having found the right formulas.
But maybe these problems have been solved by now?
Greetings,
David Matiskella
That was Micheals point. If you simplify the environment enough ( by removing
light sources, eliminating reflectors and partially transparent objects, ...) you
can simulate it. For example an old Apple II could simulate a cave with no lights
really well. :)
Graeme Humphries
Jorrit Tyberghein (Jorrit.T...@uz.kuleuven.ac.be) wrote:
: Is this true? I believe that there are still some areas which are difficult
: to render photorealisticly (I believe it had to do with water, snow and such).
: I thought it was a matter of not having found the right formulas.
: But maybe these problems have been solved by now?
The big problems that I can see lie with refraction (water, glass, etc.)
and many reflective objects. Refraction takes a lot of cpu power, and is
hard to do realtime. Same is true if you have many reflective objects
that reflect off each other in a scene.
Mark VandeWettering
In article <Pine.OSF.3.92.980623...@pigeon.qut.edu.au>,
John Ryland <j.ry...@qut.edu.au> wrote:
>
>On 22 Jun 1998, Defee Pawel wrote:
>> Take a look closer at those pictures, those shadows are real "soft" shadows. Shadow
>> buffer's "hard" shadows dont have much to do with real life.
In fact, a great deal of rendering has little do with real life. :)
>With shadow-buffers I'm sure you can make soft shadows. I think I read
>about it in "Advanced Animation and Rendering Techniques" by Watt and
>Watt.
@InProceedings{Reeves:1987:RAS,
author = "William T. Reeves and David H. Salesin and Robert L.
Cook",
title = "Rendering Antialiased Shadows with Depth Maps",
year = "1987",
month = jul,
volume = "21",
booktitle = "Computer Graphics (SIGGRAPH '87 Proceedings)",
pages = "283--291",
editor = "Maureen C. Stone",
conference = "held in Anaheim, California; 27 -- 31 July 1987",
}
Mark
--
Mark T. VandeWettering Telescope Information (and more)
Email: <ma...@pixar.com> http://www.idle.com/~markv/
No Code International Member #1173
Chris Thornborrow
David Matiskella wrote:
>
> That was Micheals point. If you simplify the environment enough ( by removing
> light sources, eliminating reflectors and partially transparent objects, ...) you
> can simulate it. For example an old Apple II could simulate a cave with no lights
> really well. :)
>
> Jorrit Tyberghein wrote:
>
> > Michael Chock wrote in message <6mm19l$fun$1...@client2.news.psi.net>...
> > >
> > >hugo elias wrote in message <358D35FC...@virgin.net>...
> > >[snip]
> > >>The amount of photorealism you can get depends on the amount of freedom
> > >>you're
> > >>willing to sacrifice. It's possible to get a totally photorealistic
> > >>scene on a
> > >>PC today.
> >
> > to render photorealisticly (I believe it had to do with water, snow and such).
> > I thought it was a matter of not having found the right formulas.
> >
> > But maybe these problems have been solved by now?
> >
> >
Oooh had to reply. This was a pet subject of mine (so called global
illumination) many years ago in academia. Let me assume you are talking
about real-time graphics and lets define that as the image being
re-calculated many times per second. Lets also assume that by
photo-realism you mean that it looks like the real world to the human
eye, that is, we would not be able to tell the difference between a
photograph of the real scene on the computer screen and the rendered
image (the Turin test of CGI).
OK assumptions defined, the exec summary is we are nowhere near.
Nowhere. If the goal could be equated to travelling across the universe
then we've reached the corner shop (and we might actually have to go
back for better shoes).
Some justification:
* When was the last time you saw a new computer game and said, what cool
graphics? Exactly, all the time. The point is you say what cool
graphics, not how did they film that, no wait, its computer graphics.
* When was the last time you went for a walk in the country and thought
how many polygons is that tree made of (actually some people do this but
what they are really asking is how many polys can I get away with to
make it look like a tree). The real world is not made of triangles.
OK bullshit over.
* Phong is being done in real time now. Phong is not a physically based
illumination model, it looks false, it always will. It is an
approximation.
* The frequency spectrum is not made of rgb, we need to sample at far
more frequencies to fool the eye
* True reflections cannot be done in real time
* General caustic effects cannot be done anywhere near real time
* Diffuse inter-reflections can only be done for none or very few moving
objects in real time
* Polygons need to be sub-pixel in size to fool the eye into seeing
curved surfaces - actually we could get to this within 5 years.
* Volumetric rendering of transparent phenomenon (smoke, mist, fire,
clouds, water) is in its infancy
I guess my overall point is, we approximate everything right now,
nothing is physically based, hell we even approximate approximations
("heres an algorithm which approximates phong shading but does it
faster"). Polygons have their limits, rgb has its limits, current
shading models have their limits, raytracing, radiosity and volumetric
techniques have their limits. We need to go back home, come up with a
Grand Unifying Theory and start again. Fair enough but it won't run in
real time. Actually we'll be lucky if it runs at all so then we'll all
spend 20 years coming up with approximations again. Errrr.....
OK lets just stick with polygons and rgb then, after all we've a hell of
a long way to go with those alone.
--
----------------------------------------------
- Chris Thornborrow tel: +44 1454 878438 -
- S/W Dev. Manager -
- PixelFusion Ltd A brave man dies but -
- Aztec West once, a coward dies -
- Bristol every day. -
----------------------------------------------
Acy James Stapp
Chris Thornborrow wrote:
> * The frequency spectrum is not made of rgb, we need to sample at far
> more frequencies to fool the eye
This, as you know, also implies a per-pixel convolution to convertback to a
three-signal display. I remember reading somewhere
(no ref, sorry) that 9 frequencies was sufficient to capture all of
the hue effects for most any surface.
> * True reflections cannot be done in real time
Generalized true reflections cannot be done in real time.
> * Polygons need to be sub-pixel in size to fool the eye into seeing
> curved surfaces - actually we could get to this within 5 years.
Actually, texture mapping with a non-affine map (i.e. a cubic insteadof linear) can
produce _very_ successful curved surfaces, except
along profile edges. AFAIK, hardware that renders spline surfaces
already exists.
> * Volumetric rendering of transparent phenomenon (smoke, mist, fire,
> clouds, water) is in its infancy
>
> I guess my overall point is, we approximate everything right now,
> nothing is physically based, hell we even approximate approximations
> ("heres an algorithm which approximates phong shading but does it
> faster"). Polygons have their limits, rgb has its limits, current
> shading models have their limits, raytracing, radiosity and volumetric
> techniques have their limits. We need to go back home, come up with a
> Grand Unifying Theory and start again. Fair enough but it won't run in
> real time. Actually we'll be lucky if it runs at all so then we'll all
> spend 20 years coming up with approximations again. Errrr.....
>
> OK lets just stick with polygons and rgb then, after all we've a hell of
> a long way to go with those alone.
>
Very understandable frustration, but things will always beapproximated. It's never,
ever, going to be possible to model
all of the interactions between light and molecules in a scene at
a molecular or atomic level, which would be requires to generate
_exactly_ the same scene.
Fortunately for us, the human visual system doesn't require that
kind of detail. We can use samply 9 frequencies instead of
hundreds. We can use texture and bump mapping.
Have you thought about how many polygons 80 million polys
actually are? At 30fps and a 10 depth complexity, that's
a 7.68 pixel polygon at 1600x1200. That's smaller than one
pixel at 640x480. Of course, 80 million polys may not _be_
enough if you insist upon motion blur (requires eye tracking,
as well) and depth-of-field effects. Of course, hardware to
render these effects could be produced easily.
Well, I've reached the end of my purposeless rant. Feel free
to argue/correct/spam/flame/whatever.
--
Acy James Stapp - Slam Software - Amp 3D Engine
ast...@slamsoftware.com http://www.slamsoftware.com
--
"Aluminum foil makes a nice hat. And it blocks the government's
mind-control rays." - Some policeman on "The X-Files"
Chris Thornborrow
Acy James Stapp wrote:
>
> Chris Thornborrow wrote:
>
> > * The frequency spectrum is not made of rgb, we need to sample at far
> > more frequencies to fool the eye
>
> This, as you know, also implies a per-pixel convolution to convertback to a
> three-signal display. I remember reading somewhere
> (no ref, sorry) that 9 frequencies was sufficient to capture all of
> the hue effects for most any surface.
According to Glassners Principles of Digital Image synthesis, a minimum
of 16 is required, improvements can be seen up to 25 samples. However
with high order polynomial basis functions to help reconstruct, 5
samples can be used.
>
> > * True reflections cannot be done in real time
>
> Generalized true reflections cannot be done in real time.
OK but I meant generalized. We are talking about photorealism aren't we?
This means real world, not limited game space. OK, You could define
photorealism to mean "if we built this game space in the real world it
would look exactly the same". If you did that though, I think its
reversing the problem and making a nonsense of it. The real world has
reflections, not "generalised reflections". Besides, generalised real
reflections can be done for small numbers of objects using sphere maps
in OpenGL and 7 passes per object on an Infinite Reality. However even
this is a very limited case. It might look like early attempts at
raytracing, it is not photorealism though.
> > * Polygons need to be sub-pixel in size to fool the eye into seeing
> > curved surfaces - actually we could get to this within 5 years.
>
> Actually, texture mapping with a non-affine map (i.e. a cubic insteadof linear) can
> produce _very_ successful curved surfaces, except
> along profile edges. AFAIK, hardware that renders spline surfaces
> already exists.
>
Again how many real world scenes have profile edges ? None of course. I
think you are falling into the trap most people do. Start with the real
world, not the "how much better this approximation is than the last"
mentality when talking about photorealism. Splines aren't enough. Render
running water with splines.
> [stuff deleted]
> >
>
> Very understandable frustration, but things will always beapproximated. It's never,
> ever, going to be possible to model
> all of the interactions between light and molecules in a scene at
> a molecular or atomic level, which would be requires to generate
> _exactly_ the same scene.
I agree and I'm not saying what we do these days is useless by any
means. I think CG has come a long long way. Hell, there are a dwindling
number here who remember that getting real time wireframe was a major,
flickering, achievement. What Pcs can do today is amazing a nothing
comapred to what we'll be doing in 5-10 years. All I want to stress is
don't mistake seeing "amazing graphics" for "we are close to
photorealism". We are not. The question I guess is, is it realistic
enough for my purposes?
>
> Fortunately for us, the human visual system doesn't require that
> kind of detail. We can use samply 9 frequencies instead of
> hundreds. We can use texture and bump mapping.
I work for a company producing per pixel shading chips for PCs. We can
do all the above (even higher sampling frequencies if we wish) but if
mother nature had a patent on photorealism, her lawyers wouldn't be
knocking on the door nobody would ever accuse us of photorealism.
> Have you thought about how many polygons 80 million polys
For a long time :-) Why pick on 80 million ? If its because a reality
monster can do it, remember all the latency in movement in that.
> actually are? At 30fps and a 10 depth complexity, that's
> a 7.68 pixel polygon at 1600x1200. That's smaller than one
No of polys = 1600 x 1200 x 60Hz (MINIMUM) x 2 (sub pixel polygon size)
= 230 Million polygons/s
Now assume that your depth complexity is say 2, this is close to 500
Million polygons to render at 60 fps at subpixel resolution on 1600x1200
screen.
Even if we coulds shift this data and render it, its still not the right
way to go (though I'd buy it today :-). We are picking on the easiest
problem to fix in some ways, the one we are accelerating toward quickest
anyways.
> pixel at 640x480. Of course, 80 million polys may not _be_
> enough if you insist upon motion blur (requires eye tracking,
> as well) and depth-of-field effects. Of course, hardware to
> render these effects could be produced easily.
The whole argument is fundamentally flawed I feel. Polygons are not
enough for photorealism. Even today, I could reduce my screen res to say
32x32 and get subpixel polygons. Would I get photorealism ? No.
>
> Well, I've reached the end of my purposeless rant. Feel free
> to argue/correct/spam/flame/whatever.
No flames but fundamentally I disagree with you.
>
> --
> Acy James Stapp - Slam Software - Amp 3D Engine
> ast...@slamsoftware.com http://www.slamsoftware.com
> --
> "Aluminum foil makes a nice hat. And it blocks the government's
> mind-control rays." - Some policeman on "The X-Files"
--
Acy James Stapp
Chris Thornborrow wrote:
> Acy James Stapp wrote:
> >
> > Chris Thornborrow wrote:
> >
> > > * The frequency spectrum is not made of rgb, we need to sample at far
> > > more frequencies to fool the eye
> >
> > This, as you know, also implies a per-pixel convolution to convertback to a
> > three-signal display. I remember reading somewhere
> > (no ref, sorry) that 9 frequencies was sufficient to capture all of
> > the hue effects for most any surface.
>
> According to Glassners Principles of Digital Image synthesis, a minimum
> of 16 is required, improvements can be seen up to 25 samples. However
> with high order polynomial basis functions to help reconstruct, 5
> samples can be used.
Thanks. I wasn't sure where I saw this; it's good to have a reference.
> >
> > > * True reflections cannot be done in real time
> >
> > Generalized true reflections cannot be done in real time.
>
> OK but I meant generalized. We are talking about photorealism aren't we?
> This means real world, not limited game space. OK, You could define
> photorealism to mean "if we built this game space in the real world it
> would look exactly the same". If you did that though, I think its
> reversing the problem and making a nonsense of it. The real world has
> reflections, not "generalised reflections". Besides, generalised real
> reflections can be done for small numbers of objects using sphere maps
> in OpenGL and 7 passes per object on an Infinite Reality. However even
> this is a very limited case. It might look like early attempts at
> raytracing, it is not photorealism though.
>
>
> > > * Polygons need to be sub-pixel in size to fool the eye into seeing
> > > curved surfaces - actually we could get to this within 5 years.
> >
> > Actually, texture mapping with a non-affine map (i.e. a cubic insteadof linear) can
> > produce _very_ successful curved surfaces, except
> > along profile edges. AFAIK, hardware that renders spline surfaces
> > already exists.
> >
>
> Again how many real world scenes have profile edges ? None of course. I
> think you are falling into the trap most people do. Start with the real
> world, not the "how much better this approximation is than the last"
> mentality when talking about photorealism. Splines aren't enough. Render
> running water with splines.
I guess I had two points intermixed there:
1. Higher than linear perspective mappings for texturing a cubic surface.
This could produce much smoother looking surfaces with current
technology.
2. Correct rendering of thhe boundaries of a cubic surface. This is a
bit trickier in hardware, I think (but then, you would know better than I).
> > [stuff deleted]
> > >
> >
> > Very understandable frustration, but things will always beapproximated. It's never,
> > ever, going to be possible to model
> > all of the interactions between light and molecules in a scene at
> > a molecular or atomic level, which would be requires to generate
> > _exactly_ the same scene.
>
> I agree and I'm not saying what we do these days is useless by any
> means. I think CG has come a long long way. Hell, there are a dwindling
> number here who remember that getting real time wireframe was a major,
> flickering, achievement. What Pcs can do today is amazing a nothing
> comapred to what we'll be doing in 5-10 years. All I want to stress is
> don't mistake seeing "amazing graphics" for "we are close to
> photorealism". We are not. The question I guess is, is it realistic
> enough for my purposes?
That's true. I guess then the question becomes "Is it realistic enoughto fool _people_
into thinking it's real?"
> > Fortunately for us, the human visual system doesn't require that
> > kind of detail. We can use samply 9 frequencies instead of
> > hundreds. We can use texture and bump mapping.
>
> I work for a company producing per pixel shading chips for PCs. We can
> do all the above (even higher sampling frequencies if we wish) but if
> mother nature had a patent on photorealism, her lawyers wouldn't be
> knocking on the door nobody would ever accuse us of photorealism.
>
> > Have you thought about how many polygons 80 million polys
>
> For a long time :-) Why pick on 80 million ? If its because a reality
> monster can do it, remember all the latency in movement in that.
>
> > actually are? At 30fps and a 10 depth complexity, that's
> > a 7.68 pixel polygon at 1600x1200. That's smaller than one
>
> No of polys = 1600 x 1200 x 60Hz (MINIMUM) x 2 (sub pixel polygon size)
> = 230 Million polygons/s
Well, the thread originally started with someone quoting 80 millionpolys as necessary to
achieve photorealism. I had assumd that
this was 80 million polys/second, but thinking about it, it's probably
80M/frame, which would be 4.8e9*depth polys/second.
> Now assume that your depth complexity is say 2, this is close to 500
> Million polygons to render at 60 fps at subpixel resolution on 1600x1200
> screen.
>
> Even if we coulds shift this data and render it, its still not the right
> way to go (though I'd buy it today :-).
Sure would scare the pants off of a Voodoo2 :)
> We are picking on the easiest
> problem to fix in some ways, the one we are accelerating toward quickest
> anyways.
>
> > pixel at 640x480. Of course, 80 million polys may not _be_
> > enough if you insist upon motion blur (requires eye tracking,
> > as well) and depth-of-field effects. Of course, hardware to
> > render these effects could be produced easily.
>
> The whole argument is fundamentally flawed I feel. Polygons are not
> enough for photorealism. Even today, I could reduce my screen res to say
> 32x32 and get subpixel polygons. Would I get photorealism ? No.
> >
> > Well, I've reached the end of my purposeless rant. Feel free
> > to argue/correct/spam/flame/whatever.
Of course, subpixel polygons at 1600x1200 more closely approximatethe resolution of the
human visual system.
A 32x32 resolution might be photorealistic were you holding waxed
paper six inches from your face.
> No flames but fundamentally I disagree with you.
Basically the only change available that I can see is direct renderingof solids/volumes
at a pixel level instead of rendering boundaries
between them at a surface level. My personal feeling is that
photorealism can be produced using a surface rendering scheme
achievable with current methods using future technology.
--
Acy James Stapp - Slam Software - Amp 3D Engine
ast...@slamsoftware.com http://www.slamsoftware.com
--
Every U.S. Citizen (and others) should read and think about:
http://www.house.gov/Constitution/Constitution.html
http://www.law.indiana.edu/uslawdocs/declaration.html
Then read:
http://www.cwrl.utexas.edu/~bill/e309m/3rd/shoemake/ethics_on_the_net/restrictions.html
Wouldn't the Declaration of Independence, if written today, be considered
"conspiracy to overthrow the government of the United States?"
John Ryland
On Fri, 26 Jun 1998, Acy James Stapp wrote:
>
> Basically the only change available that I can see is direct renderingof solids/volumes
> at a pixel level instead of rendering boundaries
> between them at a surface level. My personal feeling is that
> photorealism can be produced using a surface rendering scheme
> achievable with current methods using future technology.
Yeah, I would agree that we'll get closer to realism as we can afford to
move away from rendering polygons.
But think out several "real" scenes. A walk through a rain forest. You
look around and you see the trees swaying, the flowers moving, you can see
beads of water glissening. You walk up to a lookout and you can see for
miles and miles around hugh areas of forrest. You possible can see as many
as a few million trees from the look out!
Ok now render this. In the rain forest the trees are moving, and you have
to re-calculate your radiosity calculations as everything moves (aint
going to happen in real-time for at least 15 years). The droplets of
water, or the running stream need to be rendered with raytracing to model
the refractivity (real-time a long way off too). When you look out at the
forests from the lookout and see millions of trees, if each tree was
modelled with 1000s of polygons there would by lots of polys. Adaptive
*approximations* to the recursive ray-tracing depth, radiosity patch
interactions, to the polygon resolution of objects etc might be able to
reduce the calculations significantly.
I guess the way I understand how the 3D world "knows" what colours it is
in real-time, is because it is all done simultaneously (in parallel) and
instantaneously (or at the speed of light). The level of parallelism in
the 3D world is such that every particle is working in parallel. All the
particles do is add up all the recieved light/radiation and then emit
an appropriate quantity of light at various frequencies. Therefore the
calculation of light in the real world is made of a continuous calculation
of additions in massive parallel.
Clearly we won't reach real-time radiosity at molecular/atomic level (for
any reasonably sized scene) for at least 100 years. Adaptive control of
the resolution of all the calculations is how we will be able to make it
possible to have complex scenes with complex light interactions with
realistic computational power in the future.
What was the question? :)
John