Simplified Mitsuba implementation of the "Lightweight Photon Mapping" algorithm. This is not the implementation that was used in the paper. The implementation here lacks many optimizations and only works for area light sources (triangle meshes or rectangles). Not every feature is implemented, and by far not every corner case has been tested. Also, the performance is not optimized at all. Some parts (in particular the MIS computations and the usefulness estimation) are deliberately implemented in a less than optimal way. The focus was on the flexibility and readability of the code.
The implementation is based on Mitsuba 0.6. See the Mitsuba webpage for build instructions, dependencies, and other documentation.
Our extensions to the original Mitsuba source code consist of the following major components distributed across multiple files:
- An integrator implementing the vertex merging (VM) algorithm (vertex connection and merging without bidirectional connections)
- A function to compute the usefulness of a photon
- A function that implements the pixel classification heuristic
- A simple emission guiding approach using a 4D histogram for each light source in the scene
- Functions in the emitter classes to compute the inverse mapping of the position and direction sampling
All code, except for the required extensions to the emitters, can be found in src/integrators/lpm/.
The core contribution of the lightweight photon mapping paper, determining the usefulness of a photon, is implemented in the file usefulness.h. Histograms of the image contribution of (useful) photons are built and used for the importance sampling of emission in the files emission_sampler.h and emission_sampler.cpp.
The integrator itself is defined in the files lpm.h and lpm.cpp. The files path_sampler.h and path_sampler.cpp provide functionality to trace paths from either the camera or the lights into the scene. They are used in cam_trace.h, cam_trace.cpp, light_trace.h, and light_trace.cpp to compute image contributions from all path sampling techniques. The files mis.h and mis.cpp provide code to combine the estimates via multiple importance sampling (MIS). The derivation of the MIS weights can be found in the VCM paper, the method used to compute the weights is inspired by the approach described in the tech report on implementing VCM. The remaining files define datastructures and utility functions required for the implementation, like the hash grid used for photon mapping and the 4D histograms used for emission guiding.
The repo contains a modified version of the "Still Life" test scene used in the paper, it can be found in test/lpm/.
Furthermore, the same folder contains an IPython notebook that can load the emission histograms, which are dumped by the integrator after each iteration, and visualize them.