instant ray tracing back
(L) [2006/06/20] [lycium] [instant ray tracing] Wayback!any chance that a friend of keller *ahem* could get a preprint to some eager reviewers? ;)
that "bounding interval hierarchy" naming inspires much conjecture, which i would love to see...
(L) [2006/06/20] [toxie] [instant ray tracing] Wayback!I know. :(
Unfortunately the (new) evil EGSR copyright stuff forbids us to publish the paper on our website. :(
So:
a) Register to the EG and d'load the paper after next week
b) Mail me :)
_________________
what do you expect to do if you don't know what to do when you've got nothing to do?
(L) [2006/06/23] [toxie] [instant ray tracing] Wayback!Paper will be available from monday (maybe tuesday) on [LINK http://graphics.uni-ulm.de/].
=)
_________________
what do you expect to do if you don't know what to do when you've got nothing to do?
(L) [2006/06/26] [toxie] [instant ray tracing] Wayback!Live from EGSR06: [LINK http://graphics.uni-ulm.de/BIH.pdf]
_________________
what do you expect to do if you don't know what to do when you've got nothing to do?
(L) [2006/06/26] [goodbyte] [instant ray tracing] Wayback!Thank you! Now I know what to do tonight [SMILEY Wink]
(L) [2006/06/27] [beason] [instant ray tracing] Wayback!Very impressive.
BIH has problems with large triangles? Like cornell box? I see the discussion at the end, I'm just wondering. I guess you could apply a few iterations of loop subdivision with no smoothing. My question is: is that faster than not subdividing?
Also, re: "It also enables the much more efficient computation of motion blur [Kel03]." If I remember correctly, Keller suggests sampling in time using a quasi-random sequence, which changes from pixel to pixel. How do you suggest mixing BIH with this idea? Do you ray-trace static scenes which you rebuild quickly... for every pixel, or, for every time sample...? I guess I'm just a newb on how to ray-trace moving scenes in general, so I apologize in advance.
Thanks for any answers! Great work, btw. I particularly like how you bested InView in performance!
(L) [2006/06/27] [toxie] [instant ray tracing] Wayback!my interpretation of the motion blur comment: )
if you use dependent sampling (i.e. interleaved sampling in this case) you can build f.e. 32 BIH trees for 32 timesteps needed to have some rough motion blur estimation.
_________________
what do you expect to do if you don't know what to do when you've got nothing to do?
(L) [2006/06/28] [toxie] [instant ray tracing] Wayback!If you build everything (=tree) on demand, there -might- be a clever way to do this.
_________________
what do you expect to do if you don't know what to do when you've got nothing to do?
(L) [2006/07/02] [chris81] [instant ray tracing] Wayback!This time it's called H-tree
[LINK http://www.cgg.cvut.cz/~havran/ARTICLES/rt06submission.pdf]
(L) [2006/07/02] [Phantom] [instant ray tracing] Wayback!Ouch. Two out of three must hate the order in which these papers appeared.
EDIT: (after some reading)
Looks like Havran really wasn't aware of any of the other two papers. His link to the collision detection world is very interesting. The reference to Ooi's SKD tree is bad for the BIH paper, I think, as it basically means that the concept already existed in the very same form... Only thing new is thus the 'global heuristic'.
Nevertheless, for the moment I don't think I would like to try to implement Havran's algorithm. It's extremely complex, and I doubt that 'time to image' for his algo is better than Toxie's. Rendering times look better for H and AH trees than BIH (he even seems to confirm my findings (and sunflower's) that BIH is roughly 25% slower than kd/sah on average), but his gains on preprocessing (4-20 times compared to NlogN kd/sah) don't seem to be very spectacular, I believe Toxie's scheme does better than that on Buddha (<1sec compared to 30 secs on NlogN for the very best implementations? and there's even room for improvement, it's not even multi-threaded yet). 300ms for 100-200K scenes sounds like scores well below what Toxie reported.
But the main issue is the severe complexity. I need to read the paper again to grasp it at all, and perhaps it's just 'though writing', but as it looks now, it's really not something that you can do in a weekend. Toxie's compiler is that simple.
Right now I think I would rather test faster compilation strategies (need to implement the bucket approach, I'm handling individual prims right from the start now) and ways to improve the tree (I would like to try something with a coarse grid that subdivides large prims cheap) and faster rendering (Reshetov's traversal, as clarified by Carsten Benthin). I bet there will be some improvements to the sudden avalanche of schemes for dynamic scenes, so I'll wait for that.
I'm very interested in other people's findings after reading all these papers.
These are inspiring times. [SMILEY Smile]
_________________
--------------------------------------------------------------
Whatever
(L) [2006/07/02] [toxie] [instant ray tracing] Wayback!Especially nice:
"This research work has been mainly conducted since 1st November 2005 till 6th February 2006"
So that anyone in the world knows that he didn't copy it from the others. :)
Maybe we should have also noted something like that (from 1st June 2005 til xth x 2006 ;).
(Secret Note: Now one of the reviews we received after EGSR-submission definetly makes sense! Cause the reviewer was OBVIOUSLY pissed off by the paper (and also mentioned the SKD))
Have to check the details in the paper first, so more from my side in some days.
_________________
what do you expect to do if you don't know what to do when you've got nothing to do?
(L) [2006/07/04] [toxie] [instant ray tracing] Wayback!So.. I checked Havrans paper in more detail.
Here my thoughts:
As Phantom already pointed out, i think the whole thing is cool but way too complex to be used in practice without getting mad during the implementation (as it is even more complex like the SAH).
There is also one good idea for the BIH (that is also found on my todo-list, btw. ;) by using the "missing
case" of the BIH-splitplane-possibilities as found in Havran's thing (called BVH2 if i remember correctly),
so you can cutoff empty volume on both halfs at the same time, compared to only one. Only thing missing is
a -cool- trick to not disturb the standard BIH traversal code by this additional case.
Then there are a number of things that make me suspicious about the whole statistics in the paper:
1) Why is Havran using a "slower" kD tree in the statistics?! (=slower means incorrect, as he doesn't use the exact decisions on when triangles are sorted to the left/right/both as he points out in the description of the statistics)
To make it look worse than the (A)H tree?!
2) Why are the triangle-intersection numbers of the kD insanely high? (above 10 for convex scenes like bunny and buddha????!)
3) As a result of 1&2: Ridiculous fps-numbers for the kD (concerning the monster-machine he used (Athlon64 3800+)).
-> So is the (A)H slower than our kD implementations in all cases??! Or is it the case that all of the implementations presented by Havran equally suck?!
A final comment on the AH: The idea is of course nice, but as he pointed out himself it fails for all non-equal-distributed scenes (=real life scenes ;). This can even be verified by his own statistics.
So overall i'm a bit confused by the paper. There are obviously nice ideas in the paper, but i don't know at the moment how to interpret them (really good performancewise or really bad??!) or how to include them into my own system without blasting the code-complexity through the roof.
_________________
what do you expect to do if you don't know what to do when you've got nothing to do?
(L) [2006/07/09] [River] [instant ray tracing] Wayback!Is it only me or the link is not working anymore?
(L) [2006/07/09] [Phantom] [instant ray tracing] Wayback!Havran's .tz site is notoriously hard to reach at times.
_________________
--------------------------------------------------------------
Whatever
(L) [2006/07/11] [tbp] [instant ray tracing] Wayback!Still couldn't manage to nab a copy, is there someone kind enough to forward some my way?
_________________
May you live in interesting times.
[LINK https://gna.org/projects/radius/ radius] | [LINK http://ompf.org/ ompf] | [LINK http://ompf.org/wiki/ WompfKi]
(L) [2006/08/10] [River] [instant ray tracing] Wayback!Anyone know where I can get the H-Tree paper (that talks more about BVH2)? The link is dead and I cannot find it anywhere in google (except as HTML but there is no pictures, which probably helps quite a lot to understand).
(L) [2006/08/10] [lycium] [instant ray tracing] Wayback!i can forward it (didn't find your email addy in profile).
(L) [2006/08/14] [greenhybrid] [instant ray tracing] Wayback!metoo, please! (thx in forward thomas)
_________________
[LINK http://greenhybrid.net/] | [LINK mailto:root@greenhybrid.net root@greenhybrid.net]
back