Photon Mapper

CIS561: ADVANCED RENDERING, Penn Engineering - SPRING 2020

C++

OVERVIEW

The Photon Mapper is a Path Tracer model that starts with elements of Monte Carlo Path Tracing and augments the rendering model by calculating the contributions of photons energy. The photon mapper starts with a preprocessing pass where n photons are emitted from the light sources, bounced around the scene, and finally stored in a global illumination and caustics photon map. In the rendering step, we cast p x p + 1 ray samples per pixel, where p is the square root of the number of samples. The first p x p rays are used in calculating the Multiple Light Sampling contribution, and the final ray cast is used in calculating photon contributions. For accuracy, we only use the k closest photons to calculate the indirect and caustic contributions of photons. We use the k-d tree volumetric data structure to store the photons and optimize searching.

Render Time: 7 hours

Render Time: 11 hours

Render Time: 3 hours

1 million photons and and search radius of 0.3
1024x1024 pixels, 400 samples / pixel, depth 8 recursion

Progressive Build up

Here is a visualization of what each photon mapping feature adds to the overall render.

Multiple Importance Sampling

Photon Mapping Introduced

Recursive BSDF sampling for Specular

Recursive Photon Mapping for Specular

1 million photons and and search radius of 0.05
1024x1024 pixels, 400 samples / pixel, depth 8 recursion
Render Time: 1 hours

Basic Renders

Here, I show a comparison between the Full Lighting Integrator from the Monte Carlo Path Tracer and the Photon Mapper. The photon mapper takes significantly less time to create a converged image and appears less noisy. While the Full Lighting Integrator creates a more accurate render, the photon mapper creates a higher perceptual quality since the blur of the Photon Mapper is visually preferred over the noise of the Full Lighting Integrator.

Full Lighting Integrator

Photon Mapper

Full Lighting Integrator

Photon Mapper

1 million photons and and search radius of 0.3

Further Optimizations

Below are some optimizations discussed in class that I did not implement.

Cylindrical Search - This project currently uses a spherical search radius, utilizing a cylinder to search for photons would make sure that only photons on surfaces with similar surface normals will be used to contribute to each others illumination
Gaussian Blur - When calculating the photon energy contribution, the program currently calculates the average color value. Changing this to a gaussian blur would increase the influences of the photons closest to the intersection point.