Homec4science
Diffusion RADIANCE Photon Map (pmapmerge) Browse

RadiancePmap/README-photonflow.txt
be0a78356c8bpmapmerge

Last Change
Raw File

README-photonflow.txt
View Options

This is the README file for the RADIANCE photon flow implementation. This
software was developed as part of the project "Three-Dimensional Light Flow",
hosted by Tokyo University of Science under the supervision of Prof.
Nozomu Yoshizawa, and supported by the JSPS Grants-in-Aid for Scientific
Research (KAKENHI) under the grant number JP19KK0115.
$Revision: 1.3 $ $Date: 2020/11/16 22:41:09 $
INTRODUCTION
These source files implement a modified volume photon density estimate in
RADIANCE to evaluate the illuminance of a physical lightfield using photon
flow. This code is an experimental variant of the RADIANCE photon map extension
intended for testing prior to committing it to the RADIANCE CVS as part of the
official distribution. Whether this (admittedly specialised) functionality
should be made available to the RADIANCE community as a default build option
is debatable and may be decided in due course.
This code is essentially a C implementation of the prototype Python script
volPhotonIllum.py. As such, it is orders of magnitude faster and more
convenient, which comes to bear with large photon maps and many sensor
positions. One crucial difference between the RADIANCE and Python
implementations is that the former performs photon lookups in terms of a
fixed number of photons around the sensor position (nearest neighbour search),
while the latter does so in terms of a fixed radius (range search), and
therefore does not adapt to the local photon density, which is suboptimal.
The illuminance evaluated from photon flow approximates Cuttle's cubic
illuminance for a virtual sensor position in a physical light field.
Photon flow constitutes volume photons deposited by mkpmap in a participating
medium (mist); due to the interaction of a photon with the medium, it will
be deposited at multiple locations along its path; the density of these
photons per path is governed by the participating medium's extinction.
Evaluating cubic illuminance from photon flow entails performing volume
photon lookups for each of the 6 (virtual) cube faces corresponding to 3
orthogonal axes in positive and negative directions. The resulting search
volume is a hemisphere, and photons incident from the respective cube
face (identified by it is normal or axis) are filtered on the fly based on
their stored incident directions. This results in six independent hemispherical
lookups with individual radii adapted to the density of the incident photons
for each axis. Although this incurs more overhead, it was found that this
approach yields consistently better results than a single lookup for all axes
with a posteriori filtering, particularly with asymmetric illuminance.
It was also found that the illuminance scales with the extinction of the
participating medium, and is therefore corrected for the same. Generally,
smaller values yield more accurate results but incur longer photon distribution
times with mkpmap.
The photon mapping code was modified to generate a variant of the existing
volume photon map, referred to as a "lightflow photon map", which corrects
the photon flux according to the extinction during it generation with mkpmap.
When loaded with rtrace, a lightflow photon map triggers the cubic illuminance
evaluation via the function volumePhotonCubicIllum(). Unlike the existing
evaluation function for volume photons, volumePhotonDensity(), the lightflow
function does not apply the Henyey-Greenstein phase function to compute
outscattering (thus ignoring the participating medium), but instead performs
the per-axis lookups with a volume density estimate as described above.
Because photon flow not only includes photons that have been reflected off one
or more surfaces (indirect contributions), but also photons that have just been
emitted from light sources (direct contributions), volumePhotonCubicIllum() is
is called by multiDirectPmap() in pmapsrc.c, which normally calls the direct
photon map density estimate (typically only used for debugging). To prevent
overcounting, the ambient calculation is therefore disabled in photon flow
mode (see ampPmap() in pmapamb.c).
COMPILING
The photon flow code is not compiled by default as part of the RADIANCE suite,
and must be enabled at compile time by defining PMAP_PHOTONFLOW. This may be
accelerated by only (re)compiling in ray/src/rt if a RADIANCE build already
exists:
rmake OPT+=-DPMAP_PHOTONFLOW install
Note that these options are independent of the underlying data structure used
to store the photon map. The default options build the legacy kd-tree. If a
larger, out-of-core photon map is desired (to accommodate >1 billion photons,
for example), this build option must be enabled by defining PMAP_OOC:
rmake OPT+="-DPMAP_OOC -DPMAP_PHOTONFLOW" install
For details, see the RADIANCE Out-Of-Core Photon Map Technical Report at
http://dx.doi.org/10.13140/RG.2.1.2158.9363
Similar to stanard (planar) photon mapping, photon flow supports kernel
density estimates by weighting individual photon contributions by their
inverse distance to the lookup point. This filter reduces bias, and can
be enabled with the following:
rmake OPT+=-DPMAP_EPANECHNIKOV -DPMAP_OOC -DPMAP_PHOTONFLOW" install
This command enables the 3D Epanechnikov kernel (1 - (d_i)^2), where d_i
is the normalised distance between the lookup point and photon (i). The bias
reduction is moderate, however, and ineffective with excessive lookup
bandwidths.
Finally, the following checks verify that the photon flow code was
successfully built:
1. 'mkpmap -help' lists the '-apV' option to generate lightflow photon maps,
2. 'rtrace -help' lists the '-aS' option to toggle between planar and
spherical illuminance evaluation from lightflow photons.
USAGE
A lightflow photon map representing the physical lightfield is generated with
mkpmap by specifying the extinction of a global participating medium either
on the command line, or in a RADIANCE scene description using a mist material
with boundary geometry (e.g. to limit photonflow to a specific region of
interest). The participating medium must not absorb nor scatter, so that
a photon's trajectory is not altered as it traverses the medium. Therefore as
a convenience, mkpmap implicitly sets the albedo and scattering eccentricity
to 1.0, which is equivalent to specifying the options -ma 1 1 1 (no absorption)
and -mg 1 (forward scattering):
mkpmap -apV lightflow.vpm 1M [-apo port1 ... -apo portN] -apD 0.001 \
-me .1 .1 .1 scene.oct
The extinction -me is a free parameter here and determines the photon density
per path. Tests have shown that lower values improve accuracy but deposit
fewer photons per path, requiring mkpmap to emit more photon paths in order
to satisfy the given target photon count.
There is a caveat with using higher values of -me, namely that the two-pass
photon distribution algorithm used by mkpmap to achieve the target number of
photons may fail. The fraction of photons emitted in the 1st pass (prepass) is
specified with -apD (predistribution factor). Its default value of 0.25 is
tuned for "well behaved" geometry and materials using global or caustic
photons. This value may be too high for volume photons in a dense participating
medium, and the photon distribution overflows in the prepass with a "photon
map overflow" error, meaning that the target photon count was already exceeded
(technically this does not yield an incorrect solution, but it's not what the
user wanted). When using photon flow, this value may need to be significantly
reduced from its default of the prepass overflows.
Once the lightflow photon map has been generated, the RGB irradiance can
be evaluated using rtrace. This evaluation is enabled when loading the
lightflow photon map, which disables the normal ambient calculation.
Because the ambient calculation is entirely circumvented, the -ab parameter
is now irrelevant; the spherical volume photon density is evaluated immediately
at each sensor position. Each normal vector supplied with each sensor point
corresponds to the axis for each of the 6 cube faces, while the position is
(or should be!) identical for all 6 input records. Though redundant, this
parametrisation is consistent with rtrace convention.
sensors.dat:
<pos> <u>
<pos> <-u>
...
<pos> <w>
<pos> <-w>
rtrace -h -I -ap lightflow.vpm 1000 scene.oct < sensors.dat
Output:
<E_+u,r> <E_+u,g> <E_+u,b>
<E_-u,r> <E_-u,g> <E_-u,b>
...
<E_+w,r> <E_+w,g> <E_+w,b>
<E_-w,r> <E_-w,g> <E_-w,b>
In this example, the RGB irradiance E_i from photon flow is evaluated with
a bandwidth of 1000 photons for each axis i denoting one of the orthogonal
vectors {u, v, w} or its reverse.
The cubic scalar illuminance Esr can be calculated from these values in
a separate script as a postprocess. The advantage of this discrete approach is
that it affords access to the RGB components of each axis irradiance E_i, which
would be lost if the scalar illuminance were calculated in rtrace directly.
In addition, the mean spherical irradiance can be calculated with rtrace
to compare with the cubic scalar illuminance Esr. Rtrace accepts a boolean
option -aS to toggle between the planar irradiance evaluation for E_i as
above (corresponding to the default -aS-), and the mean spherical irradiance
evaluation using -aS+ (or just -aS):
rtrace -h -I -ap lightflow.vpm 1000 -aS+ scene.oct < sph_sensor.dat
Note that the normal vector specified for the sensor position is irrelevant
and ignored for this evaluation.
Finally, the lightflow photon distribution may be visualised by either dumping
it as a RADIANCE scene description, and rendering the photons as spheres, e.g.
pmapdump -n 1m lightflow.vpm | objview
or by dumping it as a point list for use in a point cloud viewer:
pmapdump -a -f -n 1m lightflow.vpm > photonFlow.xyz
In the latter example, pmapdump also outputs each photon's radiant flux
(in W) for further processing. See the pmapdump man page or the RADIANCE
photon map manual at http://dx.doi.org/10.13140/RG.2.1.4330.8405/1
for details.
CONCLUSION
Having reassessed the approach to evaluate cubic illuminance from volume
photons, revised the code accordingly, and tested it with asymmetric
illuminance distributions, I feel this release is easier to use, better
integrated into the existing RADIANCE framework, and more stable and mature.
However, the code, like the approach it implements, is still experimental and
needs some more testing in "real world" scenarious, so I welcome your feedback
in terms of its results, but also its usage.
Once we are confident with the results, we should make the code public and
upload it to the RADIANCE CVS repository. Initially, though, it would be
prudent to still leave photon flow an option for advanced users only.
Many thanks in advance for your cooperation and feedback. Happy photon flowing!
--Roland Schregle (roland.schregle@{hslu.ch, gmail.com})
c4science · Help