June 18, 2004

Bounding volume hierarchy is working. The scene on the left has 10000 spheres and is rendered at 4 samples/pixel without photon tracing. Instead, the color bleeding is caused by 1 bounce of path tracing. Render time is 2 minutes 46 seconds.

The algorithm is decribed in Fundmentals of Computer Graphics by Peter Shirley. The objects are divided by X, Y, Z axises recursively. Each bounding box is divided to left and right nodes. Each ray is checked with the parent bounding box, if it intersects, we check its left and right children recursively. If we reach a leaf node, we check intersection with the actual scene object bounded by that bounding box.

June 08, 2004

Participating media. I run into some problems on rendering dielectric objects. So the sphere shown here is not quite right. The volume is isotropic and homogenous. This is the simplest case. We can define absortion coeffection and scatter coeffections in this case. The isotropic phase function is a constant (1/(4*PI)). The direct lighting component is computed directly using ray tracing. The indirect component is rendered using photon map (direct estimate of radiance using photon map).

Rendering time is 1 hour 12 minutes. Should be faster, still work in progress.

June 08, 2004

Shinny metal balls and box. Notice the caustics reflected by the mirror like surfaces.

May 21, 2004

After some trials and errors, I found where the problems are and fixed them (thanks to Matt Pharr for clarifications)! So here is what I am fairly confident to be the "correct" image. Finally I am able to move on to something new! This image is rendered with 16 samples per pixel, 16 shadow rays and 20 indirect samples per primary ray. Render time: 2h 40min. The image in OpenEXR format is here: box42.exr

May 17, 2004

More sampling. I also implemented the XYZ conversion "correctly," but I am still not 100% sure.

This one is rendered using path tracing, 64 samples/pixel. You can see the noise but I think it's getting closer to be correct. I also found some mistake in my cosine distribution hemisphere random vectors. Fixed in later renderings.

These are closer to what I think would be correct. The images are identical except one was rendered at 1 sample/pixel, another at 9 samples/pixel.

But I am sure there are plenty improvements I can make...I am still working on it!

Somewhere in between, after some discussions with Brian and Rob, I realize some of the mistakes I made. The odd colors are due to color conversion error (exceeding RGB values). The last two images on the right are separate photon calculations for direct illumination and indirect illumination. I am only using the indirect illuniation from photon mapping. The direct illumination is still calculated by the path tracer.

These are photons! I was trying to see whether my photons are stored correctly and how the power is bouncing around. Ineresting?

I bet you have never seen the famous Cornell box in such colorful ways. Trust me, these are special effects I purposely put in. Photon map, what a wonderful thing!

One good thing about Henrik Wann Jensen's Realistic Image Synthesis Using Photon Mapping, is that the source code in the last few pages of the book actually work. All you have to do is type it in and boom, it works. Everything else in that book is just a collection of his papers binded together. Hmm..., it's a good book, although missing some details, which could be helpful if you are not one of his students.

After reading some of Peter Shirley's SIGGRAPH course notes on direct illumination, I am quite certain that I have that part nailed. What's probably missing is a good BRDF. I am not sure what BRDF I should use for the walls of the Cornell box. Assuming it's perfectly diffuse, the Lambert term is some constant. Is it 0.5, 1, or something else?

As you can see from the noise, we do have indirect illuminations, but I am not sure whether I'm computing it right. Nonetheless, the image on the right was sampled at 512 samples/pixel, took about 8558.58 seconds to render. In between, I have switched numerious clor conversion methods, various formulars for calculating illuminations. The result isn't that great.

I think I'm going for photon map now.

Disasters do happen, often. These are the wrong images I produced, the noise are due to the random walk path tracing algorithm. And the colors look odd. The problem is somewhere between Spectral to XYZ, XYZ to RGB, or both.

Notice how I have set up the Cornell box wrong, till I realize we are facing in positive directions acording to their data. The soft shadow is free with area light sources, as expected.

Then I spent at least one full day to try to improve (fix) the path tracer.

Here is a picture after I implemented spectral color. Even now (after 4 days), I still don't know whether I am implementing this right. But anyway, this picture looks OK, but I can assure you that the colors are wrong, even if it looks right. The attenuation term for the light is the inverse of the squared distance (which is physically correct). But I am calculating the direct lighting wrong. Basically I missed one of the two cosine terms and the area of the light.

Now you can tell I have not only a sphere intersection test working, but also implemented triangle intersection test. Both are the fastest methods I can find (steal). Also implemented tracing shadow rays, surprisingly simple.

This is The first image I produced with my ray tracer. Isn't it beatiful?

At this point I have a simple GUI set up, with my image IO ready. I also have a set of classes I always used for graphics projects, such as vector, point, etc. These classes are based on Ken's C++ source library except they are updated to use current C++ STL.