Transcript PMforONR - TWiki - Rochester Institute of Technology

Photon Mapping Development
LEO-15 DATA
SETS
(AVIRIS, IOPs)
HYDROLIGHT
Simulations
(LUT)
CONESUS
EXPERIMENT
Target Scenario,
Illumination,
IOPs
PHOTON
MAPPING
DEVELOPMENT
HYDROLIGHT
Simulations
Spectral
Matching
CONCENTRATION
MAPS
Deep Water
Scenarios
CONCENTRATION
MAPS
Large Scale
CONCENTRATION MAPS
PHOTON
MAPPING
Validation &
Verification
Summer 2003
RIT
GOALS
Atmospheric
Compensation
Rochester Institute of Technology
Shallow
Water
Scenarios
Small Scale
BOTTOM TYPE MAPS
BATHYMETRY MAPS
ALGORITHM
TRAINING/TESTING
Water Model
• Hydrolight works well for open ocean cases
• Littoral environment does not fit assumptions
 Monte Carlo approach being implemented
RIT
Rochester Institute of Technology
Basic Hydrolight World
MODTRAN Generated Sky
Detector
Random Surface
(Spatially uncorrelated)
Slabs of homogeneous
optical properties
Output is a
single point
Flat, constant bottom type
RIT
Rochester Institute of Technology
A More Complex World
MODTRAN Generated Sky
Detector
Object
interaction
Surface with
spatial structure
Underwater Plumes
Continuous/Arbitrary
distribution of optical
properties
Output is a
full scene
Variable, rough bottom types
RIT
Rochester Institute of Technology
Monte Carlo Approach
• Arbitrary 3-dimensional structure can be
handled using a Monte Carlo based approach
• Monte Carlo techniques are generally useful for
very specific problems
• General Monte Carlo based solutions are avoided
because they are very inefficient
• We are expanding on a CG technique called
“Photon Mapping” (Jensen 2001) which speeds
up the calculation of indirect illumination terms
RIT
Rochester Institute of Technology
A Simplified Scene
Source
Light is incident
on the surface of
the water
Transmitted light
is attenuated in
the water
RIT
Rochester Institute of Technology
Detector
Light reaches
the detector
Scattered and
reflected light
returns to the
surface
Forward Simulation
(Simple Monte Carlo Ray Tracing)
Source
Rays are
traced from a
light source
Light is randomly
absorbed/scattered
based on IOPs
RIT
Rochester Institute of Technology
Detector
Even fewer make
it to the detector
(Most don’t even have
a possibility of hitting
the detector)
Few rays make it to
the water surface
Backward Simulation
(Based on Photon Reciprocity)
Source
The number of ray
traces increases
exponentially with
the order of
multiple scattering
Many directions
are sampled at
each event
RIT
Rochester Institute of Technology
Detector
Rays are traced
from the detector
Rays are
randomly
propagated
until they hit a
light source
Compromise
(Two-Pass Solution )
Source
Detector
1st Pass:
2nd Pass:
Rays are traced
from light sources
Rays are traced
from the detector
Photon Map:
A searchable database
that stores the state of
the in-water light field
Once populated, the photon map can be reused by
every trace through the water
RIT
Rochester Institute of Technology
Photon Map Construction
(1st Pass – Pre-Processing Step)
Source
Rays are
traced from a
light source
Each photon is
stored in a K-D
binary tree (for
quick searches)
based on location
1
15
12
11
3
7
6
10
14
9
5
2
4
13
8
Light is randomly
absorbed/scattered
based on IOPs
RIT
Rochester Institute of Technology
At every absorption/
scattering event, a
“photon” is stored in
the map (location
and direction)
16
Photon Map Usage
(2nd Pass – Image Construction Step)
Source
Detector
Rays are traced
from the detector
The photon map is
searched and the
surrounding light
field information is
used to estimate
the in-scattered
radiance
RIT
Rochester Institute of Technology
Rays are
propagated
directly until
they hit a light
source
Example: Underwater Scene
Scattering
Lensing Effects
Jensen
RIT
Rochester Institute of Technology
MURI Water Model Composition
• Spectral Information
– Full spectral treatment corresponding to detector sensitivity
• Measured/Modeled IOPs
– Use the same inputs as Hydrolight
• Variance Reducing Sampling Techniques
– Faster convergence to correct values
• Modeled Wave Surface
– Generated from wave spectrum data
• Modularization
– Use of in-house ray tracer and sensor testing environment
RIT
Rochester Institute of Technology
16 bytes
Spectral Considerations
Position (12 bytes)
• Photon structure is condensed
• Number of photons used is still
very limited by memory
Direction (2 bytes)
KDTree Flag and
Spectral Info (2 bytes)
• Currently using a spectral
density technique
536,870,912 Bytes
33,554,432 Photons
Single wavelength
33,554 Photons/Wavelength
1000 wavelengths
RIT
Rochester Institute of Technology
Sampling Considerations
•
The majority of calculations in the model involve random
uniform samples on 2-manifolds
•
Uniform pseudo-random points have a very slow error
reduction rate (slow convergence)
Area Calculation
SUN
Phase Functions
RIT
Rochester Institute of Technology
Illumination Distribution
ACM SIGGRAPH 2003
•
Push to move towards stratified and quasiMonte Carlo sampling in CG community
•
Allows for the error estimate to improve faster
than a rate of 1/SQRT(N). (e.g. log(N)d/N)
RIT
Rochester Institute of Technology
Hybrid Sampling Algorithm
•
•
Combination of Stratified and Latin Hypercube algorithms
Guarantees uniformity without aliasing artifacts
Each cell contains one
sample (Stratified)
Each row/column pair
contains one sample (LHC)
Projection on 2-Manifolds
RIT
Rochester Institute of Technology
Sampling: Integration of 2D
SINC Function
• 5000 runs of 25, 625, and 15625 samples of a 2D
SINC function (continuous band of frequencies)
• Hybrid algorithm converges faster and produces
less outliers (Gaussian shaped)
0.200
2D Random Sampling
RIT
Rochester Institute of Technology
0.201
0.080
2D Hybrid Sampling
0.041
Wave Model Integration
S [m2s]
Parameterized
Wave Model
Or Measured
Wave Spectrum
U19.5 = 17.5 m/s
=
U19.5 = 15 m/s
U19.5 = 12.5 m/s
U19.5 = 10 m/s
Frequency [Hz]
Directional distribution
1D Frequency spectrum
2D Frequency spectrum
FT
RIT
Rochester Institute of Technology
Modularization Using Generic
Interfaces
DIRSIG (Detectors, 3D Models, etc.)
Radiometry
Solvers
IOP Server
Sample Generator
Photon Map
Water
Model
Air/Water Interface
Phase Functions
RIT
Rochester Institute of Technology
Photon Map Construction and
Searches in Parallel
•
Many computers can construct photon maps independently to form a
larger collective map
•
A single search query by radius can integrate contributions from each
independent map
RIT
Rochester Institute of Technology
Current Progress
• Majority of routines
are in place within
modular structure
• Currently working on
issues related to
sampling
• Preliminary validation
projected for
Fall/Winter 2003
RIT
Rochester Institute of Technology
Concluding Remarks
Development Goals
•
Provide a complex testing environment for target
detection algorithms
•
Allow for continuous improvement through a
modular interface
•
Provide generic tools that are able to solve new
problems without internal modification
•
Allow for automated generation of LUTs and
target subspaces
RIT
Rochester Institute of Technology
Questions?
RIT
Rochester Institute of Technology