Bilateral Guided Upsampling

This is not an official Google product.

Overview

This is an implementation of Bilateral Guided Upsampling as outlined in the SIGGRAPH Asia 2016 paper by Jiawen Chen, Andrew Adams, Neal Wadhwa, and Samuel W. Hasinoff.

Code structure

We include a MATLAB implementation of the slow global optimization algorithm and a Halide implementation of the fast approximation algorithm. We also provide a trivial GLSL shader for the performing slicing on the GPU. A full OpenGL demo application in on our roadmap.

We thank Elena Adams for the Parrot photo.

Build instructions (MATLAB)

1. Run MATLAB.
2. cd src/matlab
3. demo

Main driver files:

• bguFit Given a (low-resolution) input/output pair, fits an affine model.
• bguSlice Given an affine model and a (high-resolution) image, applies the model, producing a (high-resolution) result.
• testBGU Test harness that runs bguFit followed by bguSlice. Stores the results along with the passed-in ground truth into a result struct.
• showTestResults Displays the result struct as image figures.
• runOnFilesnames Run testBGU and showTestResults on filenames instead of matrices.
• demo.m Runs runOnFilenames on the Parrot example in images.

---

Build instructions (Halide, Linux and MacOS)

Our code should build and run on Windows but we have not tested it.

1. cd src/halide
2. Download a Halide distribution and unzip it such that you have a directory called src/halide/halide.
3. Install libpng and zlib. On MacOS, we used MacPorts and installed to the default location under /opt/local. If you use a different prefix location, edit Makefile and change MACOS_PREFIX_PATH appropriately.
4. make
5. Look at high_res_out.png and high_res_out_gray.png.

License

Apache 2.0.