A Compression Method For Spectral Photon Mapping

A Compression Method For Spectral Photon Mapping

Gorm Lai Department of Computer Science,

University of Aarhus and Deadline Games

Niels Jørgen Christensen, Informatics and

Mathematical Modelling,

Technical University of

Denmark

Spectral Rendering

Interesting Effects:

Dispersion Interference

Diffraction

Rayleight Scattering Phosphorescence

Flourescence

Why?

Problems of Spectral Photon Mapping

Spectral power distribution repr. as an array of N uniformly distributed point samples.

Size of the photon map

1 sample = 1 float = 4 bytes 1 SPD = N * 4 bytes

1 Photon map = M * N * 4 bytes

For a simple scene with 200000 photons, and

N=100, the photon map requires approx 200.000 *

100 * 4 = 76 MB

Solution?

Compression!

Basic idea

Centralize compression in the photon map.

1. Compress when adding photons.

2. Decompress when looking up

photons.

Photon Emission

Calculations done in the spectral space.

No compression until put in the photon map.

Calculations done in the spectral space. No loss Calculations done in the

spectral space. No loss

Calculations done in the spectral space. No loss Calculations done in the

spectral space. No loss

IMPORTANT! Emitted photons carry the entire spectrum of the light source.

Final Gathering

Uncompressed when looked up in the photon map.

Local BRDF

calculation done in the spectral

domain.

Pure spectral

calculation.

The Compressed Format

(also called the compact format)

Used for storing data in the photon map.

Has two forms:

• One for light paths of type LD*S+D (caustics).

• One for all other light paths (global).

Global Photons

COMPRESSED FORMAT = RGB!

Compression = Spectral to RGB, using CIE XYZ functions

Important! Information is lost, as we no longer know the

metamer we mapped from.

Global Photons

COMPRESSED FORMAT = RGB!

C = RGB Space

S = Spectral Domain S = Subset of S, onto which Smits’ method map.

c

Decompression = RGB to Spectral, using [Smi99]

Caustics Photons

Observations:

1. Remapped spectre no longer a top.

2. Too much info in original spectre

Solution:

Store in caustics in triplet, with same size as RGB.

RGBValue val;

val[0] = -1; // indicates special treatment val[1] = bin index of non-zero sample;

val[2] = value of non-zero sample;

Results!

Black Bodies

6500K, 100 bins 2800K, 100 bins 2800K, 4 bins

Wavelength dependent BSDFs

RGB 4 bins

Conclusion

Uses memory equally to RGB rep.

Very fast to compress and decompress (basically a matrix multiplication)

Suitable for hardware implementation

Questions?