This repository contains a few tests for the reparameterization technique described in the article Reparameterizing discontinuous integrands for differentiable rendering and implemented with Mitsuba 2.
To run the tests, make sure that you have compiled the version of Mistuba 2
that contains the plugin src/integrators/path-reparam.cpp:
git clone --branch pathreparam --recursive https://github.com/loubetg/mitsuba2
This branch has a default configuration that enables the compilation of GPU modes. Such modes are needed since all the tests rely on automatic differentiation. Mitsuba 2 can be compiled on Linux using:
cmake -GNinja -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang -DMTS_ENABLE_OPTIX=ON .
ninja
The CMake options -DMTS_OPTIX_PATH and -DCMAKE_CUDA_COMPILER can be used to specify the location of
Cuda and OptiX 6.5 on the system. Please see the
Mitsuba 2 documentation
for details about the compilation. Once Mitsuba 2 is compiled, the command
source setpath.sh
must be run in the Mitsuba directory. Then, the tests can be executed with Python, using the version of Python that has been used to compile Mitsuba 2. For instance:
python3 test_light_position.py
These tests render a scene and generate images with partial derivatives with respect to particular parameters of the scene, computed using the reparameterization technique. Similar images are also computed using finite differences for comparison.
Misc. tests:
test_light_position.py: derivatives wrt the position of the light source.test_light_position_pair.py: derivatives wrt the position of two light sources.test_object_position.py: derivatives wrt the position of a mesh.test_envmap.py: derivatives in shadows with a sun envmap.test_envmap_and_area.py: derivatives with both an envmap and a smooth area light.
From Fig. 3 in the paper:
test_fig3_direct_visiblity.py: derivatives on silhouettes.test_fig3_shadows.py: derivatives in shadows.test_fig3_glossy_reflection.py: derivatives in reflected light.test_fig3_refraction.py: derivatives in refraction.
From Fig. 6 in the paper:
test_fig6_variance.py: shows the effect of variance reduction.
From Fig. 11 in the paper:
test_fig11_corkscrew.py: test for bias at silhouettes.test_fig11_bias.py: test for bias in shadows.
These tests perform some optimizations of scene parameters from target images, based on gradient descent.
optim_colors.py: optimization of a rgb texture.optim_pose.py: optimization of the position of an object.optim_vertices.py: optimization of the vertices of an object.optim_light_position.py: optimization of the position of light source.