Efficient Continuous Relaxations for Dense CRFs - Code

This code implments the continuous relaxation based algorihms proposed in the following papers

"Efficient Continuous Relaxations for Dense CRF". Alban Desmaison, Rudy Bunel, Pushmeet Kohli, Philip H. S. Torr, and M. Pawan Kumar. ECCV, Springer, 2016.

"Efficient Linear Programming for Dense CRFs". Thalaiyasingam Ajanthan, Alban Desmaison, Rudy Bunel, Mathieu Salzmann, Philip H. S. Torr, and M. Pawan Kumar.

If you're using this code in a publication, please cite our papers.

This code is for research purposes only. If you want to use it for commercial purpose please contact us.

Contact: alban at robots.ox.ac.uk, rudy at robots.ox.ac.uk or thalaiyasingam.ajanthan at anu.edu.au

Our code is built on top of the software provided by Philipp Krähenbühl, which is downloaded from http://graphics.stanford.edu/projects/drf/.

How to compile the code

Dependencies:

• cmake
• OpenCV
• Eigen (included)
• liblbfgs (included)

Linux, Mac OS X and Windows (cygwin,wsl):

• mkdir build
• cd build
• cmake -D CMAKE_BUILD_TYPE=Release ..
• make
• cd ..

How to run the example

An example on how to use the code can be found in examples/inference.cpp. A sample image from MSRC dataset is given in data/ folder.

Example usage:

• ./examples/inference /path/to/data/2_14_s.bmp /path/to/data/2_14_s.c_unary dc-neg /path/to/results/

This runs the dc-neg mehod (with the default energy parameters) on the sample image and write the results to the results folder.

Cross validated parameters:

• The cross-validated energy paramters used in the paper are given in data/cv-results.txt