*"Light Makes Right"*

December 21, 1999

Volume 12, Number 2

Compiled by

All contents are copyright (c) 1999, all rights reserved by the individual authors

Archive locations: text version at
http://www.acm.org/tog/resources/RTNews/text/

HTML at
http://www.acm.org/tog/resources/RTNews/html/

You may also want to check out the Ray Tracing News issue guide and the ray tracing FAQ.

- Introduction
- Ray Tracing Roundup
- The Ups and Downs of Multithreaded Ray Tracing and Optimization, by John Stone
- A Summary of Octree Ray Traversal Algorithms, by Vlastimil Havran
- Additional Octree Traversal Notes, by John Spackman, Erik Jansen, and Joel Welling
- Z-buffer and Raytracing, by Sam Paik
- Ray Tracing Procedural & Parametric Surfaces, by Mark VandeWettering, Stephen Westin, and Matt Pharr
- Raytracing Gloss/Translucency, by Peter Eastman, Matt Pharr, and Stephen Westin
- Importance for Ray Tracing, by Per H. Christensen
- Good Computer Graphics Books, by Brent McPherson

Peter Shirley is writing a book on ray tracing. The title is "Realistic Ray Tracing", ISBN 1-56881-110-1, from A.K.Peters, http://www.akpeters.com, and looks to run 200 pages, $35. The book is not out yet, but appears to be near completion. I look forward to it, as it's by a master of the subject who is also an excellent teacher.

This issue of the RT News has two excellent longer pieces. Vlastimil Havran presents an overview of all research (that he could get his hands on) done on using octrees in ray tracing, then compares the more promising schemes against each other. John Stone gives his insights and guidelines for writing a parallel ray tracer on modern architectures.

I appreciate Vlastimil's work from the standpoint of attempting to bring more of an engineering approach to the study of efficiency schemes. Instead of presenting yet another variant scheme and testing it against some basic algorithm, he compares and contrasts algorithms from the literature and truly attempts to find what works and what does not. Evolving architectures can invalidate such results over time, but this practical research is a valuable starting place for anyone implementing octrees. I also admire his courage in giving his honest opinions on how significant the various papers are, something that is extremely useful but rarely done in the field of computer graphics.

John Stone's article discusses what to watch out for when attempting to parallelize code, and how to design from the ground up for an interactive ray tracer, vs. batch mode code. He also points out how changing architectures affect code in sometimes unexpected ways. Note that John has source code and related papers at his site.

Per Christensen turned a brief email he sent me some months ago into a full-fledged article on the use of importance in ray tracing (or lack thereof). This technique can result in noticeable savings (especially when transparent objects are involved) and his extension to using it per channel should make it all that more effective in saving computation time.

Of course, if you're a researcher you will probably ignore this issue of the RT News until January 12th, the SIGGRAPH 2000 deadline. It's also the deadline for another critical, life-changing event: it's the last day you can submit your Fantasy Graphics League team, http://www.realtimerendering.com/fgl/. The winner gets a lifetime subscription to the Ray Tracing News and many other valuable prizes, honors, titles, distinctions, and related privileges thereto, as soon as we can figure out what these might be.

The puzzle from last issue was: given an axis-aligned cube with corners at (0,0,0) and (1,1,1), cut it with the three planes x=y, y=z, and x=z. How many pieces is the cube cut into?

The answer: 6, not 8. What is interesting about this one is that there are any number of ways of thinking about it. One is simply to note that the three cutting planes all share the line running through (0,0,0)-(1,1,1), so only 6 pieces could be formed. Another solution is to look at the cube down this axis, i.e. view it from say (3,3,3). The cube has a hexagonal silhouette, and from this view the three cuts slice it like a pie. I like this visualization as it shows that (literally) a different view of the problem makes the answer obvious. A third solution presentation thinks about the equations: for x=z, the cube is divided into a part where x>z and one where x<z; same for the other two equations. 2*2*2 would be 8 pieces, but two pieces cannot exist: x<y<z<x and x>y>z>x, which leaves 6 pieces that can exist.

The next puzzle is one from "The Mathemagician and Pied Puzzler" by Berlekamp and Rodgers, http://www.akpeters.com/berlekamp2.html, an homage to Martin Gardner (free for download). You have a cube and you select at random three (different) corners. What is the chance that the triangle formed by these corners is acute (all angle < 90 degrees)? is a right triangle (has one angle == 90 degrees)?

back to contents

http://sgi.felk.cvut.cz/GOLEM/proposal1.html

If you have scenes in NFF, VRML 2.0, or MGF formats (or can convert to these, of course), please contact his group.

________

Global Illumination Compendium

Phil Dutre has been collecting formulae and equations relevant to global illumination and putting them in a single place. Visit his page for the current version (at this time about 23 pages long):

http://www.graphics.cornell.edu/~phil/GI/

________

36 Pentium II's on 17 servers running Linux beat a Cray using POV-Ray benchmark:

http://cnn.com/TECH/computing/9903/16/super.idg/

(thanks to Hector Yee for this story)

________

I've written a fur/hair shader (unfortunately not renderman yet) based on 'Rendering Fur with Three Dimensional Textures' by James T. Kajiya and Timothy L. Kay, California Institute of Technology, Computer Graphics, Volume 23, Number 3, July 1989:

http://dspace.dial.pipex.com/adrian.skilling/texels/texels.html

I've included code and images there.

- Adrian Ian Skilling ([email protected])

________

I recently completed my PhD at the MIT Graphics Group under the supervision of Profs. Julie Dorsey and Seth Teller. The title of my thesis is: "Radiance Interpolants for Interactive Scene Editing and Ray Tracing" The abstract is included below. The thesis is now available at:

http://www.graphics.cornell.edu/~kb/publications.html

Abstract:

Ray tracers are usually regarded as off-line rendering algorithms that are too slow for interactive use. This thesis introduces techniques to accelerate ray tracing and to support interactive editing of ray-traced scenes. These techniques should be useful in many applications, such as architectural walk-throughs, modeling, and games, and will enhance both interactive and batch rendering.

This thesis introduces radiance interpolants: radiance samples that can be used to rapidly approximate radiance with bounded approximation error. Radiance interpolants capture object-space, ray-space, image-space and temporal coherence in the radiance function. New algorithms are presented that efficiently, accurately and conservatively bound approximation error.

The interpolant ray tracer is a novel renderer that uses radiance interpolants to accelerate both primary operations of a ray tracer: shading and visibility determination. Shading is accelerated by quadrilinearly interpolating the radiance samples associated with a radiance interpolant. Determination of the visible object at each pixel is accelerated by reprojecting interpolants as the user's viewpoint changes. A fast scan-line algorithm then achieves high performance without sacrificing image quality. For a smoothly varying viewpoint, the combination of lazily sampled interpolants and reprojection substantially accelerates the ray tracer. Additionally, an efficient cache management algorithm keeps the memory footprint of the system small with negligible overhead.

The interpolant ray tracer is the first accelerated ray tracer that reconstructs radiance from sparse samples while bounding error conservatively. The system controls error by adaptively sampling at discontinuities and radiance non-linearities. Because the error introduced by interpolation does not exceed a user-specified bound, the user can trade performance for quality.

The interpolant ray tracer also supports interactive scene editing with incremental rendering; it is the first incremental ray tracer to support both object manipulation and changes to the viewpoint. A new hierarchical data structure, called the ray segment tree, tracks the dependencies of radiance interpolants on regions of world space. When the scene is edited, affected interpolants are rapidly identified and updated by traversing these ray segment trees.

- Kavita Bala ([email protected]) http://www.graphics.cornell.edu/~kb

________

A commercial product, but one that should be mentioned if only because commercial hardware dedicated to ray tracing is currently rare:

http://www.art-render.com/ - ART's RenderDrive

Algorithm details are relatively scarce, but you might glean something from their page http://www.art-render.com/technology/ar250.html. There was a paper on this at the hardware conference that took place concurrently with SIGGRAPH 99; no, I do not have a copy.

________

Announcing Steve's Object Builder Ver 1.4

Steve's Object Builder is a free script based tool that can be used to make 3D objects for ray tracing programs such as POV-Ray. The tool uses Perl (Practical Extraction and Report Language) as its script language interpreter. One nice thing about Perl is that it is easy to use and it is freely available on many platforms such as UNIX, Windows, and Macintosh. You do not have to be a expert Perl programmer to use this tool.

Features

RAW output POV-Ray smooth triangle output Moray UDO output DXF output POV scene file output Set smoothing angle for POV output Set edge angle for UDO output Swap the y and z axis option Scale the final object to a absolute size Hermite curves Bezier curves Bspline curves Catmull-rom curves Translate/Rotate/Scale functions Skin Extrude Extrude along a path Gravity Sweep Polygon triangulation Group object pieces by name L-systems Over 40 example objects

Check out my latest version at: http://www.carr.lib.md.us/~stevensl/

Steven Slegel ([email protected])

[A modeller in Perl? It's more an aid for doing procedural modeling, with a number of constructs for you to control. The basic mode of operation is to define a 2D contour and extrude it in some way. Since it's procedural, things like L-systems can be done relatively easily. - EAH]

________

I've (A.J. Chung) finally completed my Ph.D. duties and have placed the resulting thesis online:

http://www.doc.ic.ac.uk/~ajc/Papers/thesis.ps.bz2 (2.0 MB)

While this body of work does not concentrate purely on Global Illumination some interesting ideas are proposed, implemented and studied that may be of interest to researchers in this field:

1. Ray space partitions for ray casting acceleration

2. Detecting total occlusion of ray shafts in non-polygonal environments

3. Constructing smooth shading functions over arbitrary topologies

4. Robust classification of shadow regions -- umbra, penumbra and full illumination.

[to uncompress this file, get the bzip2 uncompressor from http://sourceware.cygnus.com/bzip2/ - EAH]

________

Panorama is a GNU project framework for 3D graphics production. The site has not been updated in the past few months, but the code appears to have a number of more advanced effects, including volumetric lighting and an implementation of the cellular texture basis function:

http://www.gnu.org/software/panorama/panorama.html

________

BMRT/Pixar Tidbits

We used the ray tracing functionality of BMRT (http://www.bmrt.org) on A Bug's Life, for 15 shots totalling around 30-40 seconds. Of course, the way the "ray server" works is that BMRT was only used for the refraction rays (and what you see "through" them), the rest of the scene was rendered with PRMan, as usual [http://www.bmrt.org/bmrtdoc/rayserver.html].

>Does anyone here know how BMRT's radiosity works? I hear it's the most

>accurate algorithm developed so far, apart from reverse raytracing and the

>like.

I've never heard that. BMRT just uses a fairly standard finite element, progressive refinement radiosity, then substitutes the indirect component of the radiosity pass for the ambient() illumination in the ray tracing pass. It's really not all that sophisticated.

- Larry Gritz ([email protected])

________

A number of polygon triangulators were mentioned in RTNv10n2. Here are some others (and new links for older ones):

http://www.cs.unc.edu/~manocha/CODE/GEM/chapter.html http://www.cosy.sbg.ac.at/~held/projects/triang/triang.html http://www.cs.cmu.edu/~quake/triangle.html http://www.cs.huji.ac.il/~danix/code/cdt.shar.gz

The last two are more concerned with triangulations which avoid producing narrow triangles, which processes such as finite element analysis prefer.

(links from Jeff Erickson, in comp.graphics.algorithms)

________

Radiance Online is a unique web-based, fully automated, 100% self-service rendering server, providing user friendly rendering based on the Radiance synthetic imaging system:

http://www.artifice.com/rendering/render.html

[Small renderings are free; an interesting concept, though it's not clear from the web page why you'd want to do this. The normal advantage of this sort of thing is that the server is much more powerful than your client machine, but the server available is not discussed here. - EAH]

________

If you enjoy links to new ray traced images or want to announce your own works, consider subscribing to the Raytraced Mailing List:

http://www.povlab.org/raytraced/

________

If you use POV-Ray, you might like TextureView, a free texture manager:

http://home.t-online.de/home/schmidtze/software/

________

The SIGGRAPH bibliography searcher, http://www.siggraph.org/publications/bibliography/, is a wonderful tool which has been upgraded this year. It does not provide you with lists of papers on particular subjects, though. Pointers to some human-generated lists on particular topics can be found on the ACM TOG page http://www.acm.org/pubs/tog/BibLook.html. I welcome additions to the focussed bibliographies listed here (ACM TOG can also provide such bibliographies a home, if you do not have web space).

________

Megahedron (RTNv10n2) lives on as Hypercosm:

http://www.hypercosm.com/

This renderer is interesting in that it uses the trick (presented in a technical sketch at SIGGRAPH 99 by Abe Megahed, who runs Hypercosm) of reflecting rays at the vertices of polygons and sampling the reflection direction. The reflection color is then added to the local shade for the vertex and the resulting color blended as usual using Gouraud shading. It gives a rough, soft, sometimes funky reflection at the cost of just a few rays, allowing real-time reflection approximation. The same technique can be used for rough shadows; the major criteria affecting quality is the underlying mesh and its resolution on the screen.

________

Ken Joy's On-Line Computer Graphics Notes and Geometric Modeling Notes pages have some valuable tutorials on a wide range of subjects. These are at:

http://graphics.cs.ucdavis.edu/GraphicsNotes/Graphics-Notes.html http://graphics.cs.ucdavis.edu/CAGDNotes/CAGD-Notes.html

(thanks to Craig Reynolds for pointing these out)

________

Another Java raytracer, with source:

http://www.l-shaped.freeserve.co.uk/computing/lray/

Me, I'm waiting for a ray tracer for the Palm PDA...

________

Source for a Java Z-buffer class can be found here:

http://home.wxs.nl/~ammeraal/grjava.html

________

Stomp3D is a modeler written in pure Java. Find it at http://msnhomepages.talkcity.com/RedmondAve/stomp3d/download.html

________

Worth a look is:

http://www-personal.umich.edu/~mstock/pages/builder.html

This site has some interesting 2D vortex, erosion, and other simulation demos, as well as some DEM converters and other tidbits.

back to contents

**Basic Parallelism Issues:**

It is well known that ray tracing is a highly parallelizable process. A naive ray tracer can be trivially parallelized by splitting up the image into chunks which are divided up evenly among processors, with each processor computing its own pixels entirely independently of the others. This simple approach will go a long ways in making a naive ray tracer run faster. With a more complex ray tracing system there are several trade-offs that have to be made:

- The granularity of the parallel decomposition (size of the work chunks)
has a direct effect on the amount of load imbalance among the CPUs. The
larger the chunks, the more imbalanced the CPUs will tend to be. The
smaller the chunks, the more balanced the CPUs loads will be. A "pixel
cyclic" decomposition (each CPU traces every Nth pixel in the image)
yields the best load balance.
- The granularity of the parallel decomposition also affects the spatial
coherence of consecutively traced rays on a CPU. Larger sized chunks of
work yield good ray coherence. The "pixel cyclic" decomposition yields
almost no ray coherence. Some classic ray tracing acceleration schemes
depend on ray coherence in order to function well (e.g. shadow cache)
Ray coherence has an additional side effect of promoting accesses to
already in-cache data structures, yielding better performance than random
access patterns.
- The way that computed pixels are stored in memory can have a direct effect on communications performance when gathering pixels together into a final image buffer. When a CPU's finished pixels are stored in a contiguous array, they are very easily transferred to their destination using a minimal amount of communications, and no requirement for "packing and unpacking" that data at source or destination. When groups of finished pixels are not in contiguous memory, they have to be sent using multiple communications calls, or by packing/unpacking at source and destination. When pixels are stored in a strided manner, packing and unpacking operations are necessary at both source and destination.

Since CPU speeds have been increasing substantially every year but memory bandwidth and latency have not kept apace, floating point operation counts are no longer the dominant issue in writing high performance rendering software. This is especially true since great progress has been made in development of ray tracing efficiency schemes. Modern CPUs are highly pipelined, often having tens of instructions in-flight at any given time. As a result, it is becoming more important to avoid causing CPU pipeline stalls than to reduce overall operation count. Typical causes of such stalls are memory load/store operations, and compare/branch operations.

Items 1, 2, and 3 from the above each have a relationship with the CPU issues in the previous paragraph. Balancing these issues is a major part of writing a fast ray tracer on parallel/multiprocessor computers. What combination of 1/2/3 is best depends on what goals a ray tracer is attempting to meet. They are vastly different for a real-time oriented ray tracer than they are a normal ray tracer.

**The Challenges of Multithreading:**

Although the parallelism issues involved in ray tracing are well known, the vast majority of the work done in parallelizing ray tracing as been done on distributed memory parallel computers. Shared memory parallel computers offer a number of advantages over distributed memory machines, but they also present a number of challenges which need to be overcome in order to fully capitalize on these advantages.

Some advantages of multithreading and shared memory hardware:

- 90% of the data structures in a typical ray tracer can be shared by all
threads/processors, yielding a tremendous savings in memory. Most notable
candidates for sharing are the scene database, texture maps, and
volumetric data. (Most modern distributed memory ray tracers do not
partition the scene database across nodes, due to the high communications
costs and complexity involved in passing rays or objects between nodes.
Most implement by keeping an entire copy of the scene on each node.)
- CPUs can synchronize and communicate with each other at tremendous speeds. Typically these speeds are order(s) of magnitude faster than the message passing hardware in distributed memory machines. It is easier to parallelize code to very fine granularity on most shared memory machines. This allows a ray tracer to extract parallelism that is difficult or impossible to squeeze out of a typical distributed memory machine.

- Most ray tracing algorithms were designed with sequential, non-reentrant
implementations in mind. Any algorithm can be adapted to yield correct
results in a multithreaded environment, but some perform very well, and
others perform very poorly. The main issue which determines performance
in a multithreaded environment is the way an algorithm uses memory,
especially memory which is shared by multiple threads.
The actual implementation issues vary considerably, including:

- The use of unprotected static/global data. When multiple threads
access global/static data without the use of mutual exclusion
locks, the values of these data items become scrambled. Adding
mutex locks to protect such data items can cure the scrambling
problem, but may also result in poor performance if done naively.
- The assumption that data structures will not be used in a reentrant
way. As with global data, shared data structures must also be
protected from unintentional concurrent use.
- The use of non-reentrant, or otherwise unsafe library functions. There are surprising number of commonly used library functions which are not thread-safe. Library functions can encapsulate problems such as a) and b), causing problems for code that uses them. A simple example of this is the rand() implementation provided by most C libraries. The rand() function contains internal static data, and is not thread-safe. When use concurrently by different threads, it yields unpredictable results ranging from incorrect behavior to fatal crashing.

- The use of unprotected static/global data. When multiple threads
access global/static data without the use of mutual exclusion
locks, the values of these data items become scrambled. Adding
mutex locks to protect such data items can cure the scrambling
problem, but may also result in poor performance if done naively.
- Straightforward modifications to existing data structures to allow
reentrancy (i.e. addition of mutex locks etc) typically yield performance
problems due to high lock contention etc. It often takes a slightly novel
approach to keep 90% of the benefits of the original sequential
non-reentrant code when going to a fully reentrant multithreaded
implementation.
- The layout of data structures in memory may cause "hot spotting", thrashing the machine's cache coherency hardware. Hot spotting occurs when two unrelated data structures occupy nearby memory addresses, close enough that they fall within the same cache line (most SMPs), or memory page (some NUMA machines). When the "unrelated" data structures are accessed concurrently by different CPUs, the cache lines that contain the data structures thrashes back and forth between the CPUs as writes occur to the two data structures.

**The Challenges of Real-Time Ray Tracing:**

The design and implementation of a ray tracing system capable of attaining real-time rendering rates (15 fps or more) poses several unique challenges to the implementer. Some of these challenges actually simplify other design decisions quite a bit.

- At a rate of 15 frames per second, there is no time to dynamically load
balance cpus to any great degree. By the time you figure out which
processors are doing too much work, and which ones are underutilized,
the frame is finished rendering. Since we're talking about real-time,
it's safe to assume that the next frame will be different (maybe quite
different) from the last one, so load balancing factors computed for a
previous frame probably aren't too helpful either. For real-time ray
tracing, we might as well throw out any ideas of load balancing, and
instead concentrate on reducing overhead in the code. The best
strategy for load balancing in this case is to come up with a good
static decomposition, balanced against ray coherency and other issues
mentioned previously.
- In order to achieve high frame rates, the per-frame overhead in the ray
tracer itself must be minimized. Simple things like allocating image
buffers, allocating and clearing counter arrays for grid data structures,
these all add to the overhead which is incurred on every frame.
- At 15 frames per second, there's no time for copying data around much.
The layout and format of the rendered pixels needs to be darn close to
its final form so that there's no need for various packing/unpacking of
strided pixel data from remote CPUs, and the number of messages sent
between CPUs should definitely be minimized.
- Unlike a regular non-real-time ray tracer, overhead from simple processes
in the ray tracing algorithm can actually have a direct effect on the
maximum attainable rendering speed. Such processes include pixel value
range clamping and quantization, generation of primary rays according to
camera parameters, allocating or clearing arrays to store counters for
"mailboxes" used by grid acceleration schemes, etc.
Things that would never make the Top-10 list of time consuming pieces of code in a regular ray tracer may actually show up as serious CPU-time contenders in a real-time ray tracer when rendering scenes that are simple enough to be rendered in real-time, or when using enough CPUs to render more complex scenes at such a speed. A simple way to see this effect is to render a "blank" screen. How fast can your favorite ray tracer render an empty scene with no objects and no lights? It is surprising how slow most ray tracers are before they are tuned for this. Remember, we get less than 1/15th of a second to do everything. Obviously, if a ray tracer can't effectively render a "blank" scene at a very high rate, then there's no way it can render a real scene at sufficient speed.

Once a ray tracer can render a "blank" scene at a very high rate, then the next step is to start rendering progressively more complex scenes, concentrating on finding unnecessary overhead when rendering very simple scenes. There are often bits of overhead which can be eliminated by pulling various initialization code outside of the per-frame sections of code, and moved into a separate centralized initialization section.

[For more information, publications, and John's source for a portable, high performance ray tracing system supporting MPI and threads, and with a front end that reads NFF, MGF, and AC3D, see his web site at http://www.ks.uiuc.edu/~johns/ - EAH]

back to contents

[What you are reading is a revised version of this article, something I normally do not do in the RTNews. The revision adds reviews of a few papers that the author did not have when the article was first published. - EAH]

This is the revised version of the article including some papers missing in the original issue; date of revision: 9th May, 2000.

In the previous RTNews RTNv12n1 there was an informal discussion about octrees, especially about ray traversal algorithms for octrees. Since I have devoted some effort in the past to implementing and testing three octree ray traversal algorithms, I have decided to try to review these and other octree traversal algorithms for ray tracing in this article.

There were at least 20 papers written about octrees and related stuff, in RTNews this topic was also heavily discussed. Last but one article about traversing octree that I am aware of was published more than three years ago, and (in this revised version) the last article was published in February 2000 at WSCG conference. I have decided to cite these papers chronologically as they were published in order to estimate the contribution of these papers, when they appeared.

Principally, creating an octree spatial data structure is nothing complex:

_____________

We start with the axis-aligned bounding box of the whole scene, consider this as a node of the octree at the depth zero. Subdivide it by three planes at all three axes; eight smaller bounding boxes are created, that are called octants. Now, recursion takes place for all the eight octants:

1) Find out how many objects intersect the corresponding bounding box of the octant.

2) If the number of objects is greater than a given threshold and the depth of the node in the octree is smaller than a maximum depth allowed, consider this octant as the scene and start from the beginning. Otherwise, mark this octant a leaf node, then assign it the objects intersecting with its bounding box (usually simply a linked list or an array of pointers to objects).

_____________

The issue of when to declare the node as a leaf is usually called termination criteria, with the number of objects and the depth of the node are the variables usually discussed and used.

So the octree is simple, is not it? So the question is what all the papers were about. Let us look at the articles.

[1] A. S. Glassner: {Space Subdivision for Fast Ray Tracing},IEEE CG&A, pp. 15-22, October 1984.

This is the first paper about octrees and tracing rays that is usually credited in citations. The traversal algorithm is simple: compute the point Q that is not currently inside the processed leaf-voxel, but in the neighbor voxel, and perform point-location for this point Q. The point location search always starts from the root of octree (for deeper hierarchies this may waste time). For memory representation Glassner used a hashing scheme over the octree node's unique IDs. At that time memory saving was an important issue, so storing empty nodes was avoided. This article is also considered historically the first one, that really brought significant speedup for ray tracing algorithm by a practical algorithm comparing with an O(N) one (the truth is that two octree & ray tracing article were published in 1983 in Japan by Fujimoto et al and Matsumato et al, but they were written in Japanese and thus were/are not publicly available). I implemented this traversal algorithm, results are below.

[2] F. W. Jansen: {Data structures for ray tracing}, Data Structures for Raster Graphics, pp. 58-72, June 1985.

The paper is a summary of the ray tracing algorithm and it also includes a section about spatial subdivision methods and ray traversal algorithms. It first outlines (on page 68) the recursive traversal algorithm for the hierarchy, that can be used both for BSP trees and octrees. No experimental results were presented.

[3] Kaplan: {Space Tracing: A Constant Time Ray Tracer}, SIGGRAPH'85 Tutorial on STAR in Image Synthesis, July 1985.

This paper is actually about BSP trees, but is included since they can emulate octrees in a simple way by forcing the BSP construction to finish only at the depth of a multiple of three. My opinion is that the title of the paper is rather an exaggeration. The data are organized, unlike [1], in binary trees using pointers, but the author explicitly mentions the subdivision of the box into eight children. It also describes a sequential traversal algorithm similar to [1].

[4] J. Sandor: {Octree Data Structures and Perspective Imagery}, C&G Vol. 9, No. 4, pp.393-405, 1985.

This paper deals with encoding objects into an octree, that are then stored as a double-linked list. It presents hidden surface removal based on ray casting; the objects used for experiments are terrain models and 3-D surfaces. The paper describes the possibility of rendering at different levels of detail, terminating the traversal octree based on user preference. In my opinion, it is probably one of the first papers concerning LOD for rendering, and the paper should likely be further considered as preceding height-field tracing and discrete ray tracing by Yagel et al.

[5] A. Fujimoto et al: {ARTS: Accelerated Ray-Tracing System}, IEEE CG&A, Vol 4, No. 4, 1986.

This paper is actually about uniform grids (called SEADS), but it discusses the traversal algorithm for octrees, that gave rise to uniform grids. It follows from the paper that uniform grids were invented because the 3DDA traversal algorithm for octrees was not as efficient as the authors had expected, particularly the vertical traversal algorithm for the octree. They discussed the 3DDA octree traversal in detail and compare it with the uniform grid 3DDA traversal. For a presented scene the total rendering time of the ray tracing with octrees was about 3 times longer than the one for uniform grids, when using the octree traversal algorithm using Glassner [1]. The time for a traversal step to the next voxel is for the uniform grid 3DDA 13 times faster than for a octree traversal step using Glassner. The total rendering time and the time for one traversal step is sometimes misinterpreted by papers that follow.

[6] Q. Peng et al: {A Fast Tracing Algorithm Using Space Indexing Techniques}, Proceedings of Eurographics'87, pp. 11-23, 1987.

This paper is about reduction of empty voxel encoding. It also discussed the intersection test of a triangle with a voxel. The authors claim that their traversal algorithm is faster in execution than [1][2], but they do not support it by any experimental results. In my personal view, the contribution of this paper is unclear.

[7] A. S. Glassner: {Spacetime Ray Tracing for Animation}, IEEE CG&A, pp. 60-70, 1988.

This paper touches on using a hybrid of BVH and octree together with the fourth dimension (time) when ray tracing animations. The details of octree traversal are not discussed.

[8] J. Arvo: {Linear-Time Voxel Walking for Octrees}, RTNews2, March 1988.

In this RTNews issue the author analyses the properties of the recursive octree traversal algorithm which actually was first published in [2] by Jansen. He uses BSP trees to explain the algorithm and especially for the analysis. Even if the concept of the traversal he reinvents is the recursive algorithm, the analysis is bright and it is interesting brain exercise; you can try it. The originality and properties of this traversal method are discussed by Eric Jansen in the subsequent RTNews3 and later in the RTNv5n1, also by Kelvin Sung in his Gems III article on BSP tree traversal and in the RTNv5n2.

[9] H. Samet: {Implementing Ray Tracing with Octrees and Neighbor Finding}, C&G Vol. 13, No. 4, pp.445-460, 1989.

This paper introduced a special traversal method algorithm based on tables to neighbors. Since the results of this algorithm are presented below for SPD scenes, we briefly describe it: 1) compute the intersection point P of a given ray with the axis-aligned bounding box of the scene. 2) Move the point P inside the bounding box by epsilon along the ray path 3) Find out the voxel V of the octree containing P using point location. 4) Compute the exit point Q of the ray with this voxel. 5) Move Q by epsilon outside the voxel boundary along the ray path. 6) Ascend octree hierarchy until Q lies in the node pointed at. This is performed using several intricate tables for faces, edges, and vertices. 7) Descend down the octree until a node W of a size greater or equal to V, using only tables routing. 8) If the reached node W is not a leaf node of the octree, descend further using point location. 9) Compute the intersection test of a ray with the objects in the voxel. 10) If an intersection exist lying between P and Q, computation is over. 11) Otherwise, go to step 4).

At first view this complicated algorithm is interesting in the sense it descends using point location if needed and ascends until some upper node is found. It does not precompute where this ascending will be in the hierarchy and in this way differs from recursive ray traversal. It also intelligently uses the information of where the ray exits the voxel to avoid point location when ascending and descending back to the same depth.

[10] H. Samet: {Application of Spatial Data Structures}, Addison-Wesley, Reading, Mass., 1990.

Samet's book deals with spatial data structures. It discusses in depth the use of quadtrees and octrees in image representation, display methods, and other uses. It follows his first book on this topic, which is more theoretically based: Design and analysis of Spatial Data Structures, 1989, from the same publisher. Concerning octrees and ray tracing, it briefly surveys traversal methods (sequential, recursive, and neighbor finding). It covers in detail the neighbor-finding technique, it is the revised version of paper [9]. To this day both books are an invaluable source of information on spatial data structures.

[11] J. Spackman and P. Willis: {The SMART Navigation of a Ray through an Oct-tree}, C&G, Vol. 15, No. 2, pp. 185-194, 1991.

The paper presents a ray traversal algorithm of octrees represented with nodes and pointers using fixed-point arithmetic. It decomposes the traversal to its vertical and horizontal phase. Both these phases are theoretically examined within the design of the new traversal algorithm. The paper also contains C-source code in the appendix, since the design of the traversal algorithm may be rather involved for the reader. No experimental measurements or comparison with previously published algorithms is presented, but the authors claim that their algorithm outperforms the one by Glassner[1].

[12] K. Sung: {A DDA Octree Traversal Algorithm for Ray Tracing}, Eurographics'91 proceedings, pp. 73-85, September 1991.

In this paper the author uses a method similar to Fujimoto of a virtual uniform grid over the octree. Unlike Fujimoto's approach, the traversal is based on the smallest voxel size and is actually 3DDA uniform grid traversal algorithm. In each virtual voxel it is checked whether it is a new processed voxel using a smart hashing scheme, which was for me at the first view rather hard to understand. The necessary condition of this approach is that subdivision planes for each interior node have to lie in the middle of the current octant. He compares the algorithm with octree by Glassner [1], Arvo/Jansen [8],[2], and uniform grid [5]. The results are not striking, no significant difference in timings is reported (7-17%), for scene "mount" from SPD the proposed algorithm is even slower.

[13] M. Agate, R. L. Grimsdale, and P. F. Lister: {The HERO Algorithm for Ray-Tracing Octrees}, Advances in Computer Graphics Hardware IV, pp. 61-73, Springer Verlag, 1991.

The authors present an enumeration of the octree subcells that, together with their presented traversal algorithm, is suitable for hardware implementation. The basic idea is to construct the integer mask (0-7) according to the ray direction, compute the signed distance to all three subdivision planes, and after sorting these distances determine the order of the voxels to be traversed. The paper does not cover any experimental results (for the hardware or software implementation) or give a comparison with another traversal algorithm. Anyway, the algorithm is simple to implement and thus probably efficient.

[14] B.S.S. Pradhan and A. Mukhopadkhyay: {Adaptive Cell Division For Ray Tracing}, C&G, Vol. 15, No. 4, pp. 549-522, 1991.

This paper claims in the introduction that it is the combination of the octree and the uniform grid. I have found out after careful reading that it is actually about BSP trees, but they are traversed as octrees. The authors do not cite any paper about BSP trees (Kaplan, or MacDonald & Booth, or Subramanian & Fussell) and do not compare their results experimentally with another method. It is interesting that the octree traversal algorithm that was briefly outlined very much resembles the one later discussed in detail and published by Gargantini [14].

[15] M.D.J. McNeill et al: {Performance of Space Subdivision Techniques}, CGF, No. 11, Vol. 4, pp. 213-220, 1992.

Authors from University of Sussex discuss the octree in the context of parallel execution on many processors. They deal mainly with the dynamic subdivision of the octree on the fly and show the statistics of the number of octree leaf nodes at different octree depths. They also provide their view of comparison among grids, octrees, and BSP trees to some extent using common sense, but without any experimental results. The interesting part is the use of the HERO [13] traversal algorithm for octrees that does not require the octants to have to be cubes. Surprisingly, they conclude the paper with the statement that the octree is a more efficient data structure than the BSP tree or grids.

[16] P-K. Hsiung and R. Thibadeau: {Accelerating ARTS}, Visual Computer, Vol. 8, No. 3, pp. 181-190, Springer Verlag, 1992.

This paper claims in the abstract that it presents a hybrid combination of uniform grid and octree, which is why I include this paper here. The authors discussed a method for when the node is subdivided into more than 8 voxels, namely 4^3 or 8^3 (though these should not be called octants any more). They call their method FINE-ARTS. For traversing between the voxels they used 3DDA algorithm at the corresponding depth of a node. But stop reading and think now......................................................................... This is actually the approach of recursive grid with limited resolution setting, is not it? The authors compare the results with Fujimoto's[4] octree and uniform grid. The authors do not cite the work of Jevans and Wyvill: {Adaptive voxel subdivision for ray tracing}, GI'89, pp. 164-172, 1989, so their contribution is rather questionable.

[17] E. Groeller: {Oct-tracing animation sequences}, SSCG'93, pp.96-101, 1993 (this conference has its page at http://www.uniba.sk/~sccg)

Author presents an approach based on octrees when rendering multiple frames of a scene with moving CSG primitives. He uses some global CSG trees and local instances and for presented scenes he saves up to 75% of the rendering time. The paper is about temporal coherence when the viewpoint is fixed. Author in this paper follows the work of Glassner[7].

[18] I. Gargantini and H.H. Atkinson: {Ray Tracing and Octree: Numerical Evaluation of the First Intersection}, CGF, Vol. 12, No. 4, pp. 199-210, 1993.

In this paper the authors propose the recursive traversal algorithm, that has special encoding of octants that have to be traversed. Authors claim the improvement of execution speed from 32% to 62% over the Samet traversal code [9], when the number of voxels is huge. Since we implemented this traversal algorithm, we outline it briefly. We assume the signed distance to the entry point and exit point of the octant is known. First we compute the signed distances with all three subdivision planes. We sort these three signed distances in an ascending order. Then we disregard the signed distances outside the interval given by the entry point and the exit point. Next we identify the octants by the positioning of the intersection points with subdivision planes with regard to the other planes coordinates. The traversal is based on the knowledge that in an octree at most four octants can be pierced. These are induced by four intervals given by three intersection points inside the current octant plus entry and exit point (five points along the ray path gives four intervals). The authors also discuss the robustness of the traversal algorithm when the ray pierces the neighborhood of intersection of more than one splitting plane, at worst the middle point of the octree node.

[19] R. Endl and M. Sommer: {Classification of Ray-Generators in Uniform Subdivisions and Octrees for Ray Tracing}, CGF, Vol. 13, No. 1, pp. 3-19, 1994.

In this paper the authors give a comparative study of traversal
techniques for uniform grids and octrees. The number of traversal
methods they implement and test is eleven, so it is the most extensive
study of octree traversal algorithms presented up to now. We only
enumerate these traversal algorithms (called in the paper ray
generators): **a)** Glassner[1] **b)**Peng[5] **c)**Samet[9]
**d)** Part of Samet[9] algorithm devoted to the
ascending/descending phase **e)** Sandor [4]. **f)** Rothe - in
Dissertation of University in Karlsruhe, Germany, 1991, written in
German. **g)** B. Froehlich and A. Johannsen - paper from 1988,
also written in German. **h)** Sung[12] **i)** Fujimoto[5]
**j)** Optimized traversal algorithm of Samet[9] **k)** two
Endl's methods, which is also another hybrid of Samet[9] and
Fujimoto[5], very similar to Fujimoto's octree traversal.

From the paper follows that for two test scenes the **j)**
algorithm is probably the fastest. Paper concludes with stating that
selection of the fastest acceleration method is scene dependent
(uniform grid versus octree).

[20] N. Stolte, R. Caubet: {Discrete Ray-Tracing High Resolution 3D Grids}, WSCG'95, pp. 300-312, 1995. (WSCG conference is at http://wscg.zcu.cz)

[This abstract was kindly provided by Nilo Stolte for the revised version of the summary.]

This paper is about discrete ray tracing (originally coined by Yagel et al.), when the objects are represented not by surfaces or polygons, but by voxels in a huge 3D grid (let us say at least 256x256x256).

This requires a significant amount of memory, so the authors attack the problem from the opposite direction, using octrees instead of 3D grids. The use of the octree is to reduce memory consumption and to efficiently skip empty space during ray traversal.

They provide a two-step approach with a fixed-point 3DDDA algorithm using 32 bits arithmetic (that loops without consulting cells while traversing octree empty regions), saving 50% of rendering time in comparison to Yagel's algorithm at the same resolution (256x256x256). The advantage increases at higher resolutions (512x512x512 or higher) where the 3DDDA reaches optimal times in each one of the two steps.

The two-step traversal algorithm considers the octree as divided in two halves, the first half being from root until a given level (generally the half of the total number of levels) and the second step starting from the leaves of the first step until the actual octree leaves. The traversal algorithm is optimized using bit level operations and caches the octree traversed cells in a stack to search the voxels from the position of the last visited voxel, thus profiting from the ray-voxel coherency.

[21] S. Worley and E. Haines: {Octrees and Whatnot}, RTNv8n2, May 16, 1995.

This RTNews contribution discussed some unusual tricks for octrees and z-buffering using octrees. Better to read it now than have any comments.

[22] N. Stolte, R. Caubet: {Discrete Ray-Tracing of Huge Voxel Space}, Eurographics'95 (CGF), Vol. 14, No. 3, pp. 383-394, 1995.

[This abstract was kindly provided by Nilo Stolte for the revised version of the summary.]

This paper is an improved version of [20], with some additional ideas included. The 3DDDA initialization and two-steps transition algorithms not shown in [20] are given here. A more detailed discussion of the technique is also given and a new 3DDDA algorithm (an improved version of the one in [20] but using 64 bit arithmetic) is presented. This algorithm is compared with Amanatides and Woo's algorithm in terms of number of operations, showing a better overall performance and better precision.

[23] I. Gargantini and J.H.G. Redekop: {Evaluation Exact Intersection of an Octree with Full Rays using only Radix-Sort and meet operations}, Compugraphics'95, pp. 278-284, 1995.

This paper deals with volume visualization as well. It uses the information about the viewpoint with regards to the bounding box of the whole scene, that is, 26 (3^3-1) regions. The visibility of octants is then predetermined by the priority order of octants for a given region, if rays hits the octree root node. This is then used for DDA to accelerate ray shooting, but unfortunately, no experimental results are reported. Similar approaches can be found in many visibility papers.

[24] K. Y. Whang et al: {Octree-R: an adaptive octree for efficient ray tracing}, IEEE TVCG, Vol. 1, No. 4, pp. 343-349, 1995.

This paper applies a cost function based on surface area heuristics to octree construction, precisely according to the paper of MacDonald and Booth ({Heuristics for Ray tracing Using Space Subdivision}, Visual Computer, pp. 153-165, Vol 6, No. 6, 1990). The idea is not to split the octree node to the eight voxels of the same size, but to position the plane(s) at the minimum of the cost function estimate. The cost function is computed independently at all three axes and the planes position is also selected independently. The paper does not introduce anything that is novel except the name Octree-R. It reports the results on scene gears, tetra, balls, and rings, the improvement is 4-47% concerning the intersection tests. The authors neglect the rendering times in the comparison (this improvement is 17-37%). They do not mention which traversal algorithm they use and also they do not compare the results with the original paper of MacDonald and Booth. However, the results of this approach are below.

[25] E. Reinhard, A.J.F. Kok, F.W. Jansen: {Cost Prediction in Ray Tracing}, EGWR'96, pp. 41-50, 1996.

This paper is not about octrees, but it estimates the cost of a spatial hierarchy for a given scene, when the hierarchy is built. The issue is whether it is possible to predict the performance of a spatial data structure before ray tracing is started; this issue is important, especially for huge scenes. The first half of the paper is about any hierarchy, but the octree is taken as the example in the article at the second half. It is difficult to restate the results here in short, it is an interesting paper worth reading, also containing experimental data.

[26] J. Revelles, C. Urena, M. Lastra: {An Efficient Parametric Algorithm for Octree Traversal}, pp. 212-219, WSCG'2000 conference, February 2000.

This paper presents recursive ray traversal algorithm, similar to Gargantini's[18] and HERO's[13] algorithms. It differs in the way the order of voxels to be traversed is determined; it does not require any sorting of signed distance of ray intersection with subdivision planes, but it creates the automaton where the voxel order to be traversed is encoded. The input for the automaton is the plane, where the ray leaves the voxel. This information together with the ray direction (integer mask 0-7) already gives the voxel traversal order. The paper contains the pseudocode and the comparison with Samet[9], SametCorner[19], SametNet[19], and Gargantini[18]. For three scenes the proposed algorithm has the best run-time. This paper I recommend to study, since it is the last paper published about ray octree traversal at present, and the published algorithm smartly uses the spatial information for ray traversal. The implementation is not difficult, it is much simpler than Samet's[9] one for example. As presented in the paper, it is designed for octrees with the mid-point subdivision approach, but I do not see the reason to be modified to work with Octree-R [24] subdivision. In my opinion, at the end the authors present unfair comparison with BSP trees based on the measurements results with one-type of regular scene (spheres in the grid), that is suitable for regular structures. They present the questionable opinion that octrees with full or nearly full leaves are more efficient than BSP trees.

_____________

I have not included papers which are specifically about BSP trees/k-d trees, since these are something rather different. The BSP tree can emulate the octree, but the octree cannot emulate the BSP tree. The cost for traversing this simulated octree using a BSP traversal algorithm can be higher or lower, but it very much depends on the implementation. I do not included on this list the papers which I have not read (mostly several master theses and technical reports), since they have been unavailable. These probably do not contain new stuff. I also omitted several papers which use octrees, but in which the octree is not a topic discussed in the paper.

From of all the papers listed above we can observe these important issues:

a) scheme for storing octree (hash table, linear scheme, node with pointers) b) construction algorithm for octree (in the midpoint of the current node, the position with best cost estimate). c) traversal algorithm (sequential, recursive, neighbor-finding etc.)

These three issues are usually somewhat intertwined. In the fall term of 1998 I decided to form my own opinion about octrees based on the implementation in the GOLEM rendering system developed at our department (http://www.cgg.cvut.cz/GOLEM). I therefore involved three undergraduate students in the fifth year for one term. Since I had only subset of the cited papers available at that time, after studying them I selected the following traversal algorithms to be implemented:

OCTREE84-C) Glassner's classic paper [1], but implemented without a hashing table, it is, using nodes with eight pointers to descendants.

OCTREE89-C) the paper of Samet[9], since the traversal code is intricate and thus appealing.

OCTREE93-C) the paper of Gargantini[18], since it was the latest one about the traversal algorithm for octrees, and generally usable for ray tracing.

OCTREE84-A) the approach of Octree-R[24] applied to [1].

The octrees are built by subdividing the center of the box, except OCTREE84-A). My strong request to students was that the generated octrees had to be the same for all the source code, so the octree can be fairly compared based on ray tracing SPD scenes. The termination criteria was the maximum depth (also called level in some papers) of the node in the octree and the number of primitives in the node. For testing below I decided to take the maximum depth as variable 4,5,6, or 7. The node is not further subdivided when the number of objects is smaller than two.

The students did their job, but later when I studied their source codes, I have found out that the coding itself is rather inefficient. Therefore, I completely rewrote their source codes, which took me about one man-month of programming, debugging, and testing. I improved the timing about three times in the best case, and thus I believe the source codes are now efficient enough. I also decided to include Octree-R for the paper of Gargantini OCTREE93-C), this is marked as:

OCTREE93-A) the approach of Octree-R applied to [18].

Letter C means center subdivision, letter A means adaptive, that is, cost estimate subdivision. The finding of the position of the splitting planes with minimum cost was performed exactly according to [20]: for the object and spatial medians and for N splitting planes equidistantly placed between these two medians, select the one with the minimum cost estimation (I used N=10). I would like to note that such a selection scheme does not always find the minimum cost estimate, since there can be a plane with better cost estimate (even a plane outside the median interval).

Here, we present only the timings for ray tracing SPD scenes, for maximum depth allowed 4,5,6, and 7. The additional parameters (number of nodes, references to objects etc.) can be found at SPD site (http://www.acm.org/pubs/tog/resources/SPD/overview.html). The parameters reported are the same as for previous comparison paper about hierarchical grids, so the comparison between grids and octrees can be performed directly. Rendering times are in seconds, testing was conducted on Intel Pentium II, 350 MHz, Linux, egcs-1.1.2, optimization switches -O2.

Method Maximum depth allowed 4 5 6 7 Scene OCTREE84-C balls 626.60 227.70 121.80 102.30 fluid 218.00 61.68 30.50 22.85 gears 316.10 247.00 260.60 339.50 lattice 64.53 56.37 61.37 65.65 mount 38.70 32.40 34.76 41.19 rings 179.80 113.50 111.80 133.30 teapot 50.55 33.17 30.72 34.12 tetra 7.27 6.34 6.79 8.00 tree 2907.00 1537.00 805.30 338.20 Scene OCTREE84-A balls 45.74 35.40 36.88 42.18 fluid 25.31 20.74 21.88 24.29 gears 206.60 209.70 246.10 330.00 lattice 60.03 49.32 51.75 52.90 mount 32.49 28.34 30.23 36.33 rings 109.30 86.47 92.81 119.40 teapot 34.77 27.65 28.56 33.27 tetra 6.48 5.31 5.05 5.57 tree 94.44 44.67 35.33 34.84 Scene OCTREE89-C balls 627.00 206.30 97.29 67.47 fluid 222.40 60.33 28.34 18.80 gears 302.70 229.00 236.30 303.40 lattice 60.72 45.22 47.68 50.58 mount 35.96 28.27 28.90 32.07 rings 173.70 101.80 94.38 107.60 teapot 46.44 27.16 22.97 23.92 tetra 6.60 5.23 5.24 5.90 tree 2887.00 1510.00 769.70 288.50 Scene OCTREE93-C balls 624.90 218.60 116.20 91.43 fluid 223.50 61.33 29.87 22.00 gears 314.10 246.10 256.60 327.90 lattice 65.89 55.45 59.00 62.19 mount 39.72 32.71 34.42 39.12 rings 181.10 112.10 108.80 126.30 teapot 48.92 30.82 27.78 29.89 tetra 6.93 5.83 6.10 6.96 tree 2893.00 1501.00 785.80 317.60 Scene OCTREE93-A balls 47.05 33.80 34.99 38.86 fluid 24.51 19.57 20.16 21.74 gears 210.90 212.00 245.50 323.60 lattice 62.33 50.15 51.39 52.36 mount 33.76 29.24 30.51 35.16 rings 111.30 87.69 92.60 115.90 teapot 33.05 25.33 25.67 29.27 tetra 6.21 5.09 4.94 5.60 tree 91.05 40.57 30.88 29.90 Best timings achieved ===================== Method/ OCTREE84-C OCTREE84-A OCTREE89 OCTREE93-C OCTREE93-A Scene balls 102.30 35.40 67.47 91.43 *33.80* fluid 22.85 20.74 *18.80* 22.00 19.57 gears 247.00 *206.60* 229.00 246.10 210.90 lattice 56.37 49.32 *45.22* 55.45 50.15 mount 32.40 28.34 *28.27* 32.71 29.24 rings 111.80 *86.47* 94.38 108.80 87.69 teapot 30.72 30.72 *22.97* 27.78 25.33 tetra 6.34 5.05 5.23 5.83 *4.94* tree 338.20 34.84 288.50 317.60 *29.90*

[I have put *'s around the fastest timing for each scene. -EAH]

We can observe some interesting properties from the statistics:

- When using subdivision for center and for cost estimate subdivision, there
is local minimum for the maximum depth allowed, let us call it d_min. This
minimum is more apparent for center subdivision, the times for depths
d_min-1 and d_min+1 are relatively higher than for adaptive subdivision.
For some scenes (tree) the local minimum was not achieved for tested
maximum depth. This property was reported by several papers, starting with
Fujimoto[4]. The optimum maximum depth allowed always depends on the number
of objects and their distribution in the scene.
- The center subdivision always shows worse performance than cost estimate
subdivision for SPD scenes for the same maximum depth allowed. On the other
hand, the preprocessing time for cost estimate subdivision is usually
higher than for center subdivision.
- We cannot predict which of the subdivision methods will create a smaller
number of leaves. Usually, for densely occupied scenes (such as "balls" and
"gears") the cost estimate subdivision creates a smaller number of leaves,
and surprisingly, still outperforms center subdivision. For sparsely
occupied scenes (such as "tree") cost estimate subdivision creates a
significantly higher number of nodes and outperforms center subdivision
considerably. My opinion is that cost estimate subdivision adapts better to
the objects' distribution in the scene.
- The best timings do not differ too much for center subdivision and cost
estimate subdivision for some densely populated scenes. But for the same
depth, cost estimate subdivision is the indisputable winner. For sparsely
occupied scenes the cost estimate subdivision is of significantly better
performance.
- The three different traversal algorithms show different number of traversal steps per ray, OCTREE84-C/A has always the largest number of traversal steps reported. It is interesting that the difference in timings between traversal algorithms for the same scene and subdivision approach do not differ too much! The reason for that is the traversal step for OCTREE84-C/A is of low computational cost, only three comparisons of coordinates per one step. Surprisingly, for center subdivision the quickest implementation is Samet OCTREE89/C. We suppose that the traversal tables are probably stored in the cache, so a traversal using mostly these tables can be also very quick. This does not correspond with [14] too much; OCTREE93 traversal cost per one step is rather high. Current hardware eliminates to some extent the performance dissimilarities for different traversal approaches. I can only state that I, as the programmer, did my best.

It is interesting to compare hierarchical grids with octrees. The performance of hierarchical grids and octrees constructed with cost estimate subdivision are very similar for SPD scenes! It is also valid for number of intersection tests per ray and number of traversal steps. The number of cells (interior nodes and leaves) generated by hierarchical grids is remarkably higher. It is caused by the construction of grids; when creating a grid subdivision, the number of cells is required to be near the number of objects in the grid bounding box. We should realize that the octree is only a special case of hierarchical grids with a specialized traversal algorithm. For SPD scenes measured and the parameters of statistics we can conclude that the hierarchies created by hierarchical grids and octrees are very similar and exhibit similar performances. Hierarchical grids are not so sensitive to the setting of parameters for grid construction as the setting of the termination criteria for octrees. It only supports the conclusion that these termination criteria for octrees, in spite of their intuitive and common sense, are not the most suitable for octree construction. It should be interesting to implement the termination criteria using the cost estimate as outlined by the article:

K.R. Subramanian and D.S. Fussell: {Automatic Termination Criteria for Ray Tracing Hierarchies}, pp. 93-100, GI'91.

The article only discusses the possibility of using such termination criteria, but no exact termination criteria are given and no experimental results were given in the article, though the cost of the hierarchy is elaborated well.

It will be also interesting to compare octrees and hierarchical/uniform grid for huge scenes containing millions of polygons, since for such scenes hierarchical grids can have a real appeal. (I do not have available such scenes if you can provide any for this purpose, please, contact me by e-mail).

What about the best efficiency scheme? It is related, but completely another problem. Now, we would like only to point out an article where the authors formally proved that the lower bound of worst time complexity of ray shooting is O(log N), where N is the number of objects. Unfortunately, these ray-shooting algorithms working in logarithmic time are proved to have at least O(N^4) space and preprocessing complexity in the worst case. See, please, the paper by L.S. Kalos and G. Marton: {Analysis and Construction of Worst-Case Optimal Ray Shooting Algorithms}, C&G, Vol. 22, No. 2, pp. 167-174, 1998 (or another paper written by them {Worst-Case Versus Average Case Complexity of Ray-Shooting}, Computing, Vol. 51, No.2, pp. 103-131, 1998). Most of the methods developed by computer graphics researches do not take worst-case complexity into account, but average case complexity is rather the main issue, even if not explicitly mentioned in the paper. It still holds that the methods aiming at the worst-case complexity that were developed by computational geometers exhibit practically unacceptable space complexity and preprocessing time complexity for common size scenes. At the present, I can only agree with the opinion that no practical acceleration technique has been proven to be prevalent for ray shooting problem for scenes with different number of objects and their distribution in the scene.

Any comments are welcome, write to me at: [email protected]

[Full statistics for the test runs are linked to from the SPD web site: http://www.acm.org/tog/resources/SPD/overview.html. - EAH]

back to contents

I saw the "Octree Traversal and the Best Efficiency Scheme" thread in RTv12n2, as spawned by the question:

"I'm trying to find out about fast methods for traversing octrees. It strikes me that you could use some kind of integer math method like a 3D Bresenham algorithm to get a list of cells a ray intersects".

As you point out Erik, my SMART algorithm adopts this approach; (though as I recall, Agate, Grimsdale and Lister's HERO approach is a variant on SMART rather than vice versa; I visited them at the University of Sussex in 1989).

Anyway, I thought you might be interested in some code I knocked up back at Edinburgh University for the lazy evaluation 3-D projections of Quaternion Julia Sets, ray traced with SMART in integer arithmetic; see the attachment. This allows for some fast voxel addressing schemes during traversal. (The code only does orthographic views at the moment, but perspective would be easy enough ... )

[I can send this code on to interested parties. -EAH]

____

Erik Jansen ([email protected]) replies:

Thanks! Good to hear some more history. My answer to Ben was just a direct reaction. I did not check on the historical correct order. HERO and SMART just came to my mind as octree versions of the ARVO algorithm. Thanks for your code.

BTW, I found the C&G article extremely difficult to read. It took me quite some time to find out that indeed the SMART algorithm was a top-down recursive algorithm. It was well hidden in the code. Looking backwards we all can see how we failed to sell our ideas in a clear and effective way.

____

John Spackman replies:

Ah, that'll be because I found it extremely difficult to write .... ;^)

I'm certainly more to blame than anyone for any confusion about the algorithm. Now that the WWW is here, I've uploaded some old slides which may clarify the approach; if you're interested, they're at http://homepages.tcp.co.uk/~john-mandy/smart/tree.ps

________

Joel Welling ([email protected]) writes:

I just noticed your discussion in the latest Ray Tracing News (RTNv12n1). The original topic was a Bresenham-like octree traversal algorithm. Do you know about Spackman and Willis' "SMART" algorithm? (The Smart Navigation of a Ray Through an Oct-Tree by Spackman and Willis, Computers and Graphics Vol. 15, No. 2, pp. 185-194, 1991). It's quite elegant in C; there is a C++ version which is perhaps less simple but more general included in my VFleet volume renderer (http://www.psc.edu/Packages/VFleet_Home). It's probably not a huge win for ray tracing, as it comes from the days before cache coherence was the big issue, but it certainly fits the description of being an integer-based octree traversal algorithm.

back to contents

Nicolas Delalondre wrote: I'm trying ways to speed up my raytracing engine ( It uses currently a box automatic hierarchy). I've heard about the use of an modified Z-buffer can do it for primary rays. Is there someone to explain me how to use and implement such a Z-buffer? Can we use it in addition to bounding boxes?

____

Almost any rendering system can be used to do the eye rays. Two major variants:

- render, but instead of a color per pixel, use an object ID per pixel. The object ID will be the first intersecting object. After "rendering" the scene, you go back over the buffer and compute the scene where the eye ray for each pixel is intersected with only one object, the one whose ID is in the object ID buffer. This extends to light sources, you can go and compute the scene from each light source's point of view (assuming you are talking about a point light) and when you go to compute the ray tree, intersection rays to the light can be looked up (approximated, really, due to the sampling grid) instead of a full traversal of the object database.
- render, then use the z-buffer by discarding objects during the database traversals that could not have an intersection with the depth in the z-buffer. Same idea for the light sources--here it is called a shadow depth buffer--and the usage is much easier than for the object ID buffer--just compare the distance to the light with the correct pixel in the shadow depth buffer, if it is further away then you conclude the point is in shadow, if it is approximately equal or closer (due to sampling grid errors) it is not shadowed by some other object.

You can see that a good hybrid solution is to use an object ID buffer for the eye rays and a shadow depth buffer for the light rays.

[I can comment further on method (a), having used Weghorst & Hooper's
original code for it and implemented the algorithm again a few times over the
years. This method is mostly a fine idea, but getting a *perfect* match
between the Z-buffer and ray tracer is impossible. If the Z-buffer says
there's a particular object at a pixel and the ray tracer doesn't hit it, no
big deal, you just shoot a full eye ray. But you can get the opposite case:
the Z-buffer says there's some object (or no object), but a full eye ray
traced through the scene actually hits a closer object, while also hitting
the Z-buffer's object further on. This is indetectable (short of shooting the
full ray) and results in an incorrect pixel. Usually not a serious problem,
but something to be aware of. One way to ameliorate the effect is to sample
an area of the Z-buffer, e.g. get the IDs of a 2x2 or 3x3 area of the
Z-buffer and test all these. -EAH]

back to contents

Anis Ahmad ([email protected]) wrote:

What techniques are available for raytracing procedural/parametric surfaces? Tessellate to hell and find the intersection between the ray in question and the generated polygons?

____

It is usually preferable to stop just before the Stygian Gate, but yes, you are headed in the right direction. :-)

I believe tessellation is in most cases the appropriate algorithm to choose. There are some refinements of course: tessellate only to the appropriate level based upon how large the patch is on the screen (to avoid linear artifacts in the image) and how strongly it is curved (to avoid missing highlights). You might also think about doing this tessellation in a lazy fashion (only tessellate when the first ray penetrates the bounding box of the parametric surface) and caching (saving and throwing away tessellations in a LRU fashion). The Toro crew at Stanford had some good ideas on how to make this work with their memory coherent raytracing research.

It is hard to argue that any numerical ray-patch intersector has any advantages. If you'd like to try implementing one, the Nishita-Sederberg Bezier clipping algorithm is one to try (described in Siggraph 90 proceedings if memory serves) or Toth's 85 Siggraph paper on interval techniques. If you are clever, I believe you can combine ideas from the two techniques to develop one which might be superior to either.

The main problem with any technique like this is that the work done by each
algorithm amounts to essentially a patch split per iteration, usually with
several iterations per ray. If a patch is hit many times by rays, you keep
paying for patch splits. Any attempt to cache the split caches is very
similar in spirit to just tessellating *anyway*, so why not just admit it up
front, and make that work well....

____

Stephen Westin replies:

Or, to put it another way, tessellation on demand with LRU caching is very similar in spirit to numerical iteration for direct intersection :).

Seriously, finding an optimal tessellation for a parametric surface is a significant computational task, and for a trimmed surface, it hurts just to think about.

____

Matt Pharr writes:

Hmm. I don't think that it's *that* terrible to find a good tessellation rate
for a parametric surface. Can you elaborate on the problems you're thinking
of?

Granted, It's probably the hardest part about handling them in general (vs. bounding and evaluation, say), but the basic "project the control points onto the screen and use them to bound parametric rate of change approach" is pretty inexpensive computationally and works quite well.

As for trimming, and ray traced patches, I've always been a fan of figuring out if the candidate hit point is actually trimmed out after an intersection with the non-trimmed patch is found, which is a (comparatively) easy 2d ray-tracing problem.

____

Stephen Westin replies:

If you try hard to find the optimal tessellation (i.e. the minimum set of polygons that represent the surface to within a given tolerance), you wind up with some pretty funky mesh topology.

> Granted, It's probably the hardest part about handling them in general

> (vs. bounding and evaluation, say), but the basic "project the control

> points onto the screen and use them to bound parametric rate of change

> approach" is pretty inexpensive computationally and works quite well.

Isn't that a conservative approach that finds the max tessellation rate needed and tessellates the whole patch to that level? I can imagine that it could over-tessellate pretty badly (which matters on large models) and possibly lead to cracking between patches.

How does that work for secondary bounces, where there is no "screen" to project onto? Or, to put it better, the mapping to the screen is more complex and possibly ill-behaved (e.g. discontinuous).

[I agree with Steve here; it's harder to tessellate well for bounce rays, and over-tessellation for the view is a serious problem, not a weird pathological case. I remember walking through a scene with a few NURBS spheres on a workstation which had NURBS support in microcode. At one point I thought the workstation had hung, but in fact it just happened to be close to one of the spheres and its image-based tessellation criterion caused it to generate millions of polygons. -EAH]

> As for trimming, and ray traced patches, I've always been a fan of figuring

> out if the candidate hit point is actually trimmed out after an

> intersection with the non-trimmed patch is found, which is a

> (comparatively) easy 2d ray-tracing problem.

Assuming that relatively little is trimmed off, that's a reasonable approach. How does that work out with real-world CAD models? It's certainly possible that, say, 90% of the surface area is trimmed off, which would be a big efficiency hit.

____

and a last tidbit from Stephen Westin:

As for procedurally-defined surfaces, the only reference that comes to mind is Kajiya's paper, "New techniques for ray tracing procedurally defined objects". It was presented at SIGGRAPH 83, and published in Computer Graphics v.17 #3. Besides laying out a method to calculate robust bounding volumes for fractal surfaces, he describes a really odd way to ray trace surfaces of revolution, involving bending space and using curved rays.

back to contents

I'm writing a ray tracer, and I want to implement gloss and translucency effects. In principle, this is simple - just use a random distribution to perturb the reflected rays about the pure specular direction, and same for the transmitted rays. As soon as I get into the details, though, it's not at all clear how to do this. I've gone through every book and article I can get my hands on, and none of them give any details, so I'm hoping someone on here can help me out.

So to put it simply: what's the best way of finding the perturbed ray directions? For consistency with the Phong shading model, I presumably need something that at least approximates a (cos)^n distribution? Also, it needs to guarantee that no ray ever gets perturbed too far, so that a "reflected" ray actually comes out the other side of the surface. And finally, to make sure the rays are uniformly distributed, it would be nice if the full distribution can be broken up into bins, and I can specify which bin a given ray should be in.

Simple, right? :)

Any help would be *very* much appreciated!

____

Matt Pharr ([email protected]) replies:

A good place for starters is the book "Monte Carlo Methods: Volume I: Basics", by Kalos and Whitlock. Every hour you spend reading that will more than make up for itself in writing a good distribution ray tracer. Andrew Glassner's "Principles of Digital Image Synthesis" may also have some good material.

In general, one wants to have an importance sampling function that maps two random numbers (u1, u2) (both between 0 and 1), to a direction on the hemisphere, given the outgoing direction. That importance sampling function should return both a new direction as well as the probability density function of choosing that angle (hello Kalos and Whitlock). The pdf roughly gives the probability of sampling that direction rather than some other direction.

Now, the Monte Carlo estimate of the integral (which is what you're trying to compute) is:

f_est = \Sum_0^n f(x_i) / pdf(x_i).

That is, it's a sum over some number $n$ of samples of the function f() at point x_i, each one weighted by 1. / pdf(x_i). In this case your f() is equal to the value of the reflection function (e.g. Phong) times the incident light in direction x_i (trace a ray and recurse to compute that...)

A simple importance sampling function just picks random directions on the hemisphere. This is a good importance function for a diffuse surface (as you might imagine), but it's lousy for a very glossy surface (since most of the directions that it chooses will be in directions where the reflection function is low).

How to turn (u1,u2) into a direction on the hemisphere? One option is to ignore (u1,u2) and pick random vectors in the unit square until you find one that's inside the unit sphere (exercise: show that you get a non-uniform distribution of directions if you just randomly pick a direction in the unit square.)

Vector3 sampleHemisphere() { Vector3 w; do { w.x = RandomDouble(-1., 1.); w.y = RandomDouble(-1., 1.); w.z = RandomDouble(0., 1.); } while ( w.length_squared() > 1. || w.z == 0.); return w.hat(); // normalize w }

In this case, the pdf is 1./2pi. (K&W is your friend).

A better option is to map (u1,u2) to a direction. See Pete Shirley's web page for some techniques (I think these are in some Graphics Gems book): http://www.cs.utah.edu/~shirley/. Glassner probably discusses these, too. I just looked at Shirley's web page and was reminded of some good references: see in particular "Distribution Ray Tracing: Theory and Practice".

If you have a more interesting reflection function (e.g. Phong), you're right about approximating the cos^n stuff. For Phong, it turns out that you can sample it exactly.

What you want to do is use u1 to compute an offset from the reflected direction R (dtheta) and then u2 to compute a rotation around that offset (phi). These and some vector geometry give you the new direction.

Computing dtheta (and the pdf) is the tricky part. The following function importance samples the Phong BRDF. I really urge you to save this message, read K&W and maybe some of Glassner and try to derive the importance sampling function on your own. Really, you'll be glad you did.

// returns pdf and fills in *incoming. pdf is zero if *incoming is // below the horizon, in which case no ray should be traced and this // sample's value should be zero. double brdfSampleR(const Vector3 &outgoing, Vector3 *incoming, double sample[2], double exponent) { double costheta = pow(sample[0], 1. / (exponent+1)); double sintheta = sqrt(1 - costheta*costheta); double phi = sample[1] * 2. * M_PI; Vector3 R = reflection(outgoing); // compute perfect reflection direction Vector3 Ru, Rv; R.coordSystem(&Ru, &Rv); // computes two vectors Ru and Rv to form // a little coordinate system around R. *incoming = R * costheta + sintheta * (sin(phi) * Ru + cos(phi) * Rv); if (dot(*incoming, N) <= 0.) return 0.; else return (exponent + 1) * pow(costheta, exponent); }

The next fun thing to do with a distribution ray tracer is area lights; see "Monte Carlo Methods for Direct Lighting Calculations", again on Peter Shirley's papers web page.

Hope this helps.

____

Stephen Westin (westin*[email protected]) responds to Matt's posting:

> A better option is to map (u1,u2) to a direction. See Pete Shirley's web

> page for some techniques (I think these are in some Graphics Gems book):

Gems III, pp. 80-83.

> If you have a more interesting reflection function (e.g. Phong), you're

> right about approximating the cos^n stuff. For Phong, it turns out that

> you can sample it exactly.

And also for Eric Lafortune's multi-cosine-lobe representation, which has at least a chance of physical accuracy. For most glossy surfaces, it's probably your best bet. A RenderMan shader implementing the function (but not random sampling of it) is available at http://www.graphics.cornell.edu/~westin/lafortune/lafortune.html.

____

Peter Eastman replies:

Thanks a lot to both of you for your suggestions! I just glanced over Shirley's paper on distribution ray tracing, and it definitely looks like something I should spend some time with. Lafortune's model looks rather more elaborate than anything I was figuring on trying to do, but who knows - maybe I'll go crazy and give it a shot. I'll take a look at the other references you suggested too.

I was thinking more about the problem last night, and came up with another idea for how to handle it. My idea was:

- Find a unit vector in the ideal specular direction.
- Add a random vector to it, chosen from a uniform spherical distribution. I'd choose it using basically the method you posted - pick random x, y, and z values, and throw them out if the length is too large. The radius of the distribution (less than 1) determines how much reflections are blurred.
- Renormalize the resulting vector, and use it as the reflected ray direction.

Does this sound reasonable? It's certainly not "accurate" in any sense of the word, but it should be fast and look reasonably good.

One other question for Matt. In the code you posted, I noticed that when your reflected ray direction ends up on the wrong side of the surface, you just return 0. Doesn't this cause reflections to get dimmer at grazing angles? Or do you repeatedly call the routine until it gives you a nonzero value? I was figuring that when this happens, I would just flip the reflection vector to the other size of the surface, so that the brdf would get squashed, but keep the same total intensity.

Thanks a lot!

____

Matt Pharr replies:

> Does this sound reasonable? It's certainly not "accurate" in any sense of

> the word, but it should be fast and look reasonably good.

That would definitely work, in the sense of giving you pictures, but I think it would cause weird artifacts. As I understand the idea, you're effectively sampling uniformly over a cone around the reflection direction. Because the value of the effective reflection function doesn't fall off on the edges of the cone, the reflections might look weird. One possibility is to probabilistically reject rays, rejecting them more often the farther they get from the R vector. This is known as a rejection sampling technique.

> One other question for Matt. In the code you posted, I noticed that when

> your reflected ray direction ends up on the wrong side of the surface, you

> just return 0. Doesn't this cause reflections to get dimmer at grazing

> angles?

You would call the function n times, n chosen ahead of time. For any times that the reflected ray is below the horizon, you just assign zero to that ray's result and still divide the sum of the samples by n when you're done. That flip trick, though done in certain widely used renderers, is wrong. An intuition is that you happen to have chosen to sample the function somewhere that it's zero. You'd like to avoid doing that, but when you do, that's relevant information and you should record it. As such, you take your lumps and record a zero result for the direction that you sampled.

Another way of thinking about it: consider uniform sampling of a constant function over the hemisphere. You take n samples. Each sample has pdf 1/2pi (see previous message). Say the function is 2 everywhere. Your estimate of the function is 1/n * Sum_n (2 / (1./2pi)) = n/n * 4pi = 4pi. Now say that you're foolish and you pick random directions over the sphere. The pdf is 1/4pi (surface area of unit sphere). The function is 2 for half of the directions you choose (on average) and 0 for the other half, making an average value of 1. Your estimate should be 1/n * Sum_n (0+2)*.5/(1/4pi) = n/n * 1./(1/4pi) = 4pi. If instead you reflected all of those rays to go in the up direction, you'd incorrectly get 8pi as an answer.

There are a couple of different ways to think about sampling in a distribution ray tracer. One, which seems to be where you're starting from (and which is how Cook et al described the process originally), is to find a way to generate samples that mimic the distribution you're interested in. In effect, it's necessary to integrate the function and invert the integral to do this. Sometimes this isn't possible.

Another way of thinking is in terms of importance sampling: you've got some function that is maybe equal to the function that you're trying to integrate or maybe just similar to it in some way. You figure out how to generate samples from that function's distribution. Then, when estimating the final value of the integral, you divide by a normalization factor that accounts for the distribution that you actually sampled from. This is a more general way of going about all this stuff, though it is more tricky and adds another layer or two to get right. But I'd urge you to try to head in that direction at some point in the future..

____

Peter Eastman responds:

> That would definitely work, in the sense of giving you pictures, but I

> think it would cause weird artifacts. As I understand the idea, you're

> effectively sampling uniformly over a cone around the reflection direction.

> Because the value of the effective reflection function doesn't fall off on

> the edges of the cone, the reflections might look weird.

Actually, it isn't uniform since the displacement vector can lie anywhere inside the sphere, not just on the surface. So the probability is maximum at the center of the cone (where the sphere is thickest), and goes to zero at the edges.

> You would call the function n times, n chosen ahead of time. For any times

> that the reflected ray is below the horizon, you just assign zero to that

> ray's result and still divide the sum of the samples by n when you're done.

Ah, I understand. So at each surface, you fire off many reflection rays to sample the distribution? I'm following Cook's method of only firing one ray, and picking it according to the desired distribution. In this case, flipping the ray isn't really wrong, it just means that your rays have a different distribution at glancing angles.

> There are a couple of different ways to think about sampling in a

> distribution ray tracer. One, which seems to be where you're starting from

> (and which is how Cook et al described the process originally), is to find

> a way to generate samples that mimic the distribution you're interested

> in. In effect, it's necessary to integrate the function and invert the

> integral to do this. Sometimes this isn't possible.

Shirley's Distribution Ray Tracing paper tells how to do that for a Phong distribution, so I can use his equations. Thanks again for telling me about that paper!

back to contents

**1. Uses of importance in ray tracing**

[Hall83] describes how one can keep the fraction of contribution with each ray and use that to prune the ray tree -- if the result of tracing a ray is known a priori to contribute very little to the final image, there is no need to trace it. But this fraction (which is the same as importance) can also be used to speed up other parts of ray tracing:

- Reducing the sampling of area light sources (for soft shadows).
- Reducing the probability of sampling distant point lights (a la [Shirley96]).
- Reducing the number of rays in distribution ray tracing for glossy reflection.
- Reducing the computational effort spent in procedural texture shaders.
- Reducing the effort to calculate normals (bump maps can be simplified or skipped, geometric normals can replace interpolated normals).
- Simplifying intersection tests if there is a level-of-detail representation of geometry.

There are probably several more uses than these. In addition, there are also some uses that are not strictly "traditional" ray tracing:

- If the constant "ambient" term is replaced by a better approximation of soft indirect illumination, then the computational effort spent on estimating that indirect illumination can be reduced.
- Increasing the step length in ray marching for volumes with participating media.

**2. Examples**

If a part of a scene is seen through a semi-transparent object that lets 10% of the light through, the importance of the objects behind it is not small enough that we can avoid tracing transparency rays, but the illumination calculations can be simplified drastically using some or all of the simplifications listed above.

In a scene containing water, glass, ice or similar refractive materials, each ray intersection will usually spawn both a reflection and a refraction ray. For multiple layers of these materials, there is an exponential explosion in the number of rays to be traced -- unless importance is used. The reflection coefficients depend on the angle of incidence (Fresnel's law), so some rays will have very little influence on the final image. If we keep an importance associated with each ray, we can not only avoid tracing many rays, we can also simplify calculations for many other rays that are somewhat important (not so unimportant that we can skip them altogether, but unimportant enough that we don't need to do all the calculations associated with them to full precision.)

**3. Why each color band should have separate importance**

In general, there should be a separate importance "band" for each color band. For an RGB color representation, there should be red, green, and blue importance. To see why, consider the following example: if an eye ray hits a yellow specular object, the importance of the reflection ray will have 0 blue component. If the reflection ray then hits a purely blue diffuse object illuminated by an area light source, we can skip all the illumination rays to sample that area light source since we know that none of the illumination of the blue object is going to show up in the rendered image.

If we had only a scalar importance, we would assign the reflection ray importance 1, and the illumination of the blue object would be important enough to require tracing illumination rays to the area light source.

**4. Importance transport in ray tracing**

Importance is transported like light, but emitted from the eye. This means that eye rays have importance (1, 1, 1) and that at reflections and refractions, the importance should be multiplied by the reflection/refraction coefficients to get the importance for the new ray.

Let's look at an example to see how this works in practice. If an eye ray hits a yellow ideal mirror, the reflection ray will get importance (1, 1, 0). If that reflection ray hits a 50% transmissive object, the refraction ray will have importance (0.5, 0.5, 0). If that ray hits a diffuse red surface with diffuse reflection coefficients (0.6, 0, 0), the illumination ray will have importance (0.3, 0, 0) -- assuming the light is a point light. For an area light sampled by n illumination rays, each illumination ray should have an importance smaller than (0.3, 0, 0), although probably not 1/n smaller. (Perhaps the square root of 1/n is good, I'm not sure.)

**5. Importance in global illumination**

Importance (also known as "potential" and "visual importance") is formally defined as the adjoint of radiance. It has been used in neutron transport theory since the late 1940s and was introduced to the global illumination community by Smits et al. in 1992 [Smits92] (they used it to speed up the calculation of a view-dependent radiosity solution). Since then, importance has been used to speed up calculations for many variations of both Monte Carlo and finite element methods. Please refer to the bibliography below for a list of some of the many papers about importance in global illumination.

**6. Why is importance not used more in ray tracing?**

Given that importance for ray tracing is conceptually simple, is simple to program, and gives significant speedups, the obvious question is: why don't more commercial and shareware ray tracers use it? In my opinion, it's a crying shame to burn CPU cycles computing results to an unnecessarily high accuracy.

For ray tracers without a shader interface, all the importance book-keeping can be done behind the back of the user. For ray tracers with a shader interface, shader writers need to make their shaders exploit importance. This will increase their programming effort since each shader should be able to perform both simplified (approximate) calculation and full (accurate) calculation. But the pay-off in reduced rendering time is really worth it. And if some shader writers don't want to use importance (for whatever reason), they can just ignore it.

I would be interested in a discussion on why importance is not more widely used in ray tracers. I would also like to hear some "battle stories" from people implementing importance in ray tracers. The way I see it, only the Blue Moon ray tracer has a good excuse not to use importance since it has to adhere to the Renderman specification. But what about all the other commercial and shareware ray tracers without importance?

[As a data point, the two ray tracers I developed for commercial products had Hall's importance method built in to them, turned on by default. It saves a number of nearly useless rays, and we never had complaints. That said, you usually have to be conservative in doing importance testing for a classical ray tracer. For example, you generally want to assign an importance to a particular material overall, so that you either generate all reflection rays or none of them - generating just a few of them could lead to noticeable artifacts. -EAH]

**References:**

[Hall83] Roy Hall, Donald Greenberg. "A testbed for realistic image synthesis". IEEE Computer Graphics and Applications, 3(8):10-20. November 1983. Describes the use of importance (although it wasn't called that) to avoid tracing rays whose contribution to the image is insignificant.

[Shirley96] Peter Shirley, Changyaw Wang, Kurt Zimmerman. "Monte Carlo techniques for direct lighting calculations". ACM Transactions on Graphics, 15(1):1-36. January 1996. Describes how to sample only a subset of a large number of direct light sources without introducing bias.

[Smits92]: see bibliography below.

**Annotated bibliography** (very incomplete!) of papers on importance in global
illumination:

Larry Aupperle, Pat Hanrahan. "Importance and discrete three point transport". Fourth Eurographics Workshop on Rendering, pp. 85-94. June 1993. Uses importance for a finite element method akin to hierarchical radiosity, but able to handle glossy reflections.

Philippe Bekaert, Yves Willems. "Importance-driven progressive refinement radiosity". Rendering Techniques '95 (Proceedings of the 6th Eurographics Workshop on Rendering), pp. 316-325. Springer-Verlag, 1995. Extension of progressive refinement radiosity to use importance in the refinement criterion.

Per Christensen, David Salesin, Tony DeRose. "A continuous adjoint formulation for radiance transport". Fourth Eurographics Workshop on Rendering, pp. 95-104. June 1993. Introduces exitant importance, a quantity closely related to the adjoint of radiance. Exitant importance is transported exactly like radiance, but emitted from the eye.

Per Christensen, Eric Stollnitz, David Salesin, Tony DeRose. "Global illumination of glossy environments using wavelets and importance". ACM Transactions on Graphics, 15(1):36-71. January 1996. Uses importance for a finite element method simulating glossy global illumination.

Philip Dutre, Yves Willems. "Importance-driven Monte Carlo light tracing". Fifth Eurographics Workshop on Rendering, pp. 185-194. 1995. Uses importance for photon tracing.

Philip Dutre, Phillipe Bekaert, Frank Suykens, Yves Willems. "Bidirectional radiosity". Rendering Techniques '97 (Proceedings of the 8th Eurographics Workshop on Rendering), pp. 205-216. Springer-Verlag, 1997. Radiosity method using importance, avoids computing any radiosity for unimportant patches.

Jeffrey Lewins. Importance, the Adjoint Function: the Physical Basis of Variational and Pertubation Theory in Transport and Diffusion Problems. Pergamon Press, 1965. Book with a detailed description of the use of importance in neutron transport theory. Mentions that the term "importance" was coined by Harry Soodak in 1948.

Attila Neumann, Laszlo Neumann, Philippe Bekaert, Yves Willems, Werner Purgahofer. "Importance-driven stochastic ray radiosity". Rendering Techniques '96 (Proceedings of the 7th Eurographics Workshop on Rendering), pp. 111-122. Springer-Verlag, 1996. Uses importance to modulate patch sampling probabilities in order to reduce variance in important parts of the scene.

Sumant Pattanaik, S. Mudur. "The potential equation and importance in illumination computations". Computer Graphics Forum, 12(2):131-136. Eurographics, 1993. Introduced importance for photon (particle) tracing.

Sumant Pattanaik, S. Mudur. "Adjoint equations and random walks for illumination computation". ACM Transactions on Graphics, 14(1):77-102. January 1995. Describes importance for photon (particle) tracing.

Ingmar Peter, Georg Pietrek. "Importance driven construction of photon maps". Rendering Techniques '98 (Proceedings of the 9th Eurographics Workshop on Rendering), pp. 269-280. Springer-Verlag, 1998. Traces "importance photons" from the eye position before tracing photons from the light source. Uses the importance to decide where to trace and store photons.

Peter Schr"oder and Pat Hanrahan. "Wavelet methods for radiance computations". Fifth Eurographics Workshop on Rendering, pp. 303-311. June 1994. Uses importance for adaptive refinement of wavelet radiance.

Brian Smits, James Arvo, David Salesin. "An importance-driven radiosity algorithm". Computer Graphics (Proceedings of ACM SIGGRAPH '92), pp. 273-282. July 1992. Seminal paper introducing the use of importance to global illumination; uses it to speed up hierarchical radiosity.

Laszlo Szirmay-Kalos, Balazs Csebfalvi, Werner Purgathofer. "Importance driven quasi-random walk solution of the rendering equation". Computers & Graphics, 23(2):203-212. 1999.

Eric Veach, Leonidas Guibas. "Bidirectional estimators for light transport". Fifth Eurographics Workshop on Rendering, pp. 147-162. Uses importance for bidirectional Monte Carlo simulation of global illumination.

Eric Veach. "Non-symmetric scattering in light transport algorithms". Rendering Techniques '96 (Proceedings of the 7th Eurographics Workshop on Rendering), pp. 81-90. Springer-Verlag, 1996. Shows that if the bidirectional scattering distribution function BSDF is non-symmetric, importance is scattered differently than radiance. Examples are specular refraction and the use of shading normals different than geometric normals.

back to contents

Zachary Turner wrote: can anyone recommend some good books on computer graphics theory? Obviously there's the Foley book, can anyone recommend some other good ones?

First, I would recommend anything by Alan Watt. I find his writing style very accessible and understandable. His book, "3D Computer Graphics" is more accessible to a beginner than Foley et al. IMHO. As a computer graphics programmer, I find the Graphics Gems books indispensable but they are not ideal tools for learning the theory.

Here's a few good books from my bookshelf:

- 3D Computer Graphics, Watt

- Advanced Animation & Rendering Techniques, Watt & Watt

- Computer Graphics: Principles & Practice, Foley et al.

- Mathematical Elements for Computer Graphics, Rogers/Adams

- Procedural Elements of Computer Graphics, Rogers

- Computer Graphics Handbook, Mortenson

- Computational Geometry in C, O' Rourke

- Graphics Gems I-V - The Computer Image, Watt/Policarpo

- Principles of Digital Image Synthesis I&II, Glassner

- Splines for use in CG and Geometric Modeling, Bartels/Beatty/Barsky

- Texturing and Modeling: A Procedural Approach, Ebert

- A Trip Down the Graphics Pipeline, Blinn

- Dirty Pixels, Blinn

back to contents

Eric Haines / [email protected]