Reinforcement Learning for Light Transport Rendering

This program aims to reproduce the Dahm & Keller (2017) Reinforcement Learning method (Learning Light Transport the Reinforcement Learning Way) for rendering. The method is based on the idea of using a reinforcement learning algorithm to learn how to sample light paths in a scene, which can lead to more efficient rendering.

The implementation is done in Python and uses mitsuba 3, a differentiable physically based renderer. The code is structured in a way that allows for easy experimentation and modification of the reinforcement learning algorithm.

The main components of the code include:

• A scene representation that defines the geometry, materials, and light sources in the scene.
• A reinforcement learning agent that learns to sample light paths based on the rewards received from the rendering process.
• A rendering loop that iteratively updates the agent's policy and renders the scene using the learned sampling strategy.

Dependency Installation

To run the program, you will need to have Python installed along with the necessary dependencies, including mitsuba 3.

You can install the required dependencies using pip:

pip install mitsuba

Usage

To run the program, simply execute the main script:

python render_cbox_rl.py

You should obtain four images:

• render_result: The final rendered image using the reinforcement learning method.
• render_no_update: A reference image rendered using the same algorithm but without updating the policy, serving as a baseline for comparison.
• render_no_guiding: A reference image rendered using a standard path tracing method without any guiding, serving as another baseline for comparison.
• render_result_mi: An image rendered using the multiple importance sampling (MIS) technique, which is a common method for improving the efficiency of path tracing.

Testing

Several test cases are available to validate the proper functioning of the program.

Run for example: pytest test_local_irradiance.py or python -m pytest test_local_irradiance.py

• test_local_irradiance.py: This test case checks the correctness of the local irradiance estimation.