Blockwise Multi-Order Feature Regression for Real-Time Path Tracing Reconstruction

This is the code used in the paper:

"Blockwise Multi-Order Feature Regression for Real-Time Path Tracing Reconstruction".

by Koskela M., Immonen K., Mäkitalo M., Foi A., Viitanen T., Jääskeläinen P., Kultala H., and Takala J.

Virtual reality and Graphics Architectures (VGA) group, Tampere University, http://www.tuni.fi/vga/

Datasets

The dataset to run the code is around 19GB because it contain 60 frames animations of 7 scenes with references and feature buffers all in single-precision .exr format.

The datasets can be found here: http://www.tuni.fi/vga/bmfr

Building

Make sure that you have the dataset's "inputs" folder at the location defined by INPUT_DATA_PATH, which can be found in opencl/bmfr.cpp

You need to rebuild the project if you change the files in the location without changing the INPUT_DATA_PATH. It changes the camera_matrices.h and the makefile/project does not check its modification date because the path to it is defined in the bmfr.cpp file.

Linux

Install OpenCL driver and OpenImageIO library.

make && ./bmfr in opencl folder should build and run the code.

Windows

Building and running the bmfr.sln with Visual Studio 2017 should work out of the box.

Notes

Defines in the bmfr.cpp file can be used to:

• Edit some of the BMFR algorithm's parameters
• Run the code with different inputs
• Change some of the optimizations for finding the fastest runtime on your target hardware