pouët.net

thanks for the advice chaps - how do I post a image here - i'll post a screen shot so u can see the problem

thats using the simple code on the previous pages - u can see the forshortening problem.

I found something for you. Look at this document
WarpEngine

If I will find some free time, I will try implement it - it looks very simple.

I think that it could be good technology for static locations.

awesome - just reading it now - thanks man

cool article - maybe thats a solution to the overdraw issue


create 3 depth maps from the octree using the nearest 3 visible surfaces

and then image warp for your final image - it's almost like shadow mapping e.t.c..

had a "great Idea" lol

the perspective problem is a right pain - u need the Z.

But with four 2D points u can find the perspective center quite easily by simply drawing the diagonals - the perspective center is where the lines cross. This of course needs a divide to solve the intersection (might be a short cut to exploit the fact the 2d points are square - too lazy to figure this out). So I was thinking - maybe some sort of software hologram? - 2 sets of visual 2D info combine to make a 3d image ? the octree has all the info - just gotta work out how to make the 2 sets of data.

Why are you so afraid of divides?

engine coder mate - divides/if statments give me the willies!

what's wrong with if?

conitionals are generally bad (maybe on a risc the're ok is small doses - conitional execution e.t.c)

modern PC's are pretty good at handling conditionals - but u still don't want them in your tight low level loops

branch prediction on the x and ps3 is broken - if's e.t.c.
and a L2 cache miss is 500-600 cycles

the 2 things combined are baaad :P

u can get around it by hand coding - but it's a pain.

whynot2000: 'if' sends me to branching hell on the GPU if administered in lethal doses.

Yes, the diagonals are lines in both world and screen space, so the point where they intersect is perspective correct. But it only applies to the centre of each square. I can't see how it helps you split the edges.

You're right about conditional code, of course. It's usually best to avoid it even if you get a bunch of multiplications by 1 and that sort of thing. But divisions really aren't that bad, especially with pairing and pipelining and all that.

yup it doesn't help finding the edges - it just put me into the line of thinking that the 2 2d lines combined give u 3d info. thus extending that fully - 2 sets of 2d info - to create the 3d image - ala software hologram type thing. Dunno if it's doable - in fact I don't even know where to start, just thought it was interesting as crazy as it sounds :)

Jim Blinn - Inferring Transforms (buy it there or warez it somewhere else ;)

There's a routine at the end that converts four screen space points into a 3x3 matrix for getting perspective correct homogenous coordinates (x,y,w). Works out of the box (used it for some art projection stuff on non-orthogonal surfaces), no divs/ifs, and all your perspective and edge finding problems are solved. :)

sweet thanks dood

found this too:

http://digital.library.adelaide.edu.au/dspace/bitstream/2440/48497/3/02chapters1-5.pdf

ZJ: If divisions give you the willies BECAUSE you're an engine-coder, you should find a different discipline. Seriously.

hehe thanks for the advice kusma :) but i'll stick to it all the same

3D from 2D Point Correspondences:
Fundamental Matrix

http://www.youtube.com/watch?feature=player_embedded&v=Irf-HJ4fBls

:O

O_______________________________o;

another demo from unlimited delayed technology?

Okay, the video is pretty convincing. Now give me an exe.

it's nice :)
i wonder how much TBs/PBs those datasets are tho :)

I thought the video was laughably unconvincing personally, he's got to be the worst salesman ever :D The bit where he says they're a group of scientists picking the hardest problems out there and taking them down one by one, hilarious!

The tech on the other hand looks pretty convincing. The bit that does really get me though is the disk access. If you have say a 1TB data set, and you're viewing at say 1 megapixel res, even at best case you're asking for say 3 million bytes from disk. Per frame, so multiply that by at least 10. And those bytes can come from potentially anywhere on the disk. It's going to need a very fast 1TB+ SSD. Yet they show it running well over USB 2 or a network.