DeepestScatter

Real-time photo-realistic clouds path tracer with Lorenz-Mie scattering.

This is an enchancement on paper by Disney Research Deep Scattering.

Result

Published and presented at SIGGRAPG Asia

State of the project

Preparation

• [x] Generate a sample cloud in Houdini
• [x] Exporter of volumetric data from Houdini
• [x] Importer of volumetric data for python and c++

Ground Truth

• [x] Naive implementation of path tracer that only takes into account Beer-Lambert law and direct scattering
• [x] Interactive visualisation of results
• [x] Naive implementation that also takes into account Lorenz-Mie phase function.
• [x] Implement multiple scattering using Monte-Carlo method
• [x] Implement progressive path tracing
• [x] Implement custom progressive path
• [x] Implement Reinhard tone mapping
• [x] Implement next event estimation
  

Dataset

• [x] Generate more clouds
• [x] Gather dataset of (cloud_id, point_in_cloud, sun_direction, view_direction, radiance)
• [x] Gather dataset of samples around the point according to the original paper
• [x] Separate the dataset for training, test, validation with total size of 15`000`000; leave a few clouds out of those sets for visual evaluation

Training

• [x] Build a model in pytorch as described in the original paper
• [x] Train the model
• [x] Implement an efficient neural-based renderer in c++
• [x] Compare to the original results

  

Research Experiments with different architectures.

• [x] Shadow maps
• [x] Latent space light probes
• [x] Light probe rotation
• [x] Light probe interpolation (trilinear, barycentric)
• [x] Evaluate radiance only for the outer layers of the cloud