Inverse Transport Networks

This repository is an implementation of the method described in the following paper:

"Towards Learning-based Inverse Subsurface Scattering" [project website]\ Chengqian Che, Fujun Luan, Shuang Zhao, Kavita Bala, and Ioannis Gkioulekas\ IEEE International Conference on Computational Photography (ICCP), 2020

Getting Started

These instruction constains three parts:

• ITNSceneFiles - scripts and additional file needed for rendering our dataset images and derivatives
• Renderer - a Monte-Carlo Differentiable renderer based on Mitsuba 0.5.0
• Learning - codes for training and evaluating our networks

Rendering Scripts and Dataset

We used Mitsuba to generate our dataset. Our image file name follows the convention as: [shape]_e[sunlight_direction]_d[sigmaT]_a[albedo]_g[g]_q[sampleCount].exr

For example, one can render the following scenes:

• shape: cube
• sunDirection: azimuth angle 30 and elevation angle 60
• materials: sigmaT 100, volumetric albedo 0.8 and g 0.2
• number of samples: 4096

mitsuba scenes/cube_sunsky.xml -Dmeshmodel=cube -DsigmaT=100 -Dalbedo=0.8 -Dg=0.2 -DnumSamples=4096 -Dx=0.433 -Dy=0.866 -Dz=0.25 -o cube_e30_d100_a0.8_g0.2_q4096.exr

One can also render a class of images using the following bash scripts with Sun Grid Engine:

./create_jobs_sge.sh

Differentiable Renderer

We developed our differentiable renderer based on Mitsuba 0.5.0 and it compiles the same way as compiling Mistuba. To compile and render a signle image with derivatives:

cd renderer
scons
mitsuba scenesAD/cube_sunsky.xml -Dmeshmodel=cube -DsigmaT=100 -Dalbedo=0.8 -Dg=0.2 -DnumSamples=4096 -Dx=0.433 -Dy=0.866 -Dz=0.25 -o cube_e30_d100_a0.8_g0.2_q4096.exr

The current renderer supports computing derivatives with respect to the following parameters:

• scattering parameters: extinction coefficient, volumetric albedo and g in Henyey-Greenstein phase function
• bsdf parameters:roughness, reflectance and weights in a mixure of bsdfs
• lighitng: weights in a mixture of environment maps

Learning code

Codes used to train and evaluate our approach is inside folder learning/. Pre-trained models with 5 different networks can be downloaded here.

Built With

• Pytorch
• Pandas
• scikit-image
• scikit-learn
• pyEXR
• scipy
• EdgeBox for creating weight maps

The networks were trained using Amazon EC2 clusters. All image names are in ITNSceneFiles/imgNames/. One can evaluate our model by doing:

python eval.py

And to use our models to initialize analysis by synthesis, one can run:

 python eval_calibrated.py