https://github.com/tvercaut/densecrf
GitHub archive of Philipp Krähenbühl's dense CRF code
https://github.com/tvercaut/densecrf
conditional-random-fields crf densecrf
Last synced: about 1 month ago
JSON representation
GitHub archive of Philipp Krähenbühl's dense CRF code
- Host: GitHub
- URL: https://github.com/tvercaut/densecrf
- Owner: tvercaut
- License: bsd-3-clause
- Created: 2023-07-05T21:01:46.000Z (almost 2 years ago)
- Default Branch: main
- Last Pushed: 2023-07-06T10:04:32.000Z (almost 2 years ago)
- Last Synced: 2025-01-25T18:32:23.531Z (3 months ago)
- Topics: conditional-random-fields, crf, densecrf
- Language: C++
- Homepage: http://graphics.stanford.edu/projects/drf/
- Size: 1.52 MB
- Stars: 2
- Watchers: 1
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.txt
- License: LICENSE
Awesome Lists containing this project
README
DenseCRF - Code
=============
http://graphics.stanford.edu/projects/drf/
This software pertains to the research described in the ICML 2013 paper:
Parameter Learning and Convergent Inference for Dense Random Fields, by
Philipp Krähenbühl and Vladlen Koltun
and the NIPS 2011 paper:
Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials, by
Philipp Krähenbühl and Vladlen Koltun
If you're using this code in a publication, please cite our papers.
This software is provided for research purposes, with absolutely no warranty
or suggested support, and use of it most follow the BSD license agreement, at
the top of each source file. *Please do not contact the authors for assistance
with installing, understanding or running the code.* However if you think you
have found an interesting bug, the authors would be grateful if you could pass
on the information.
Changes to the original code
----------------------------
The only major difference in this released version of the code is, that I directly
compute the gradient of the permutohedral lattice, instead of the general Gauss
Transform (3 line formula in p.6 in ICML 2013 paper). The gradient of the
permutohedral lattice evaluated the exact gradient of the approximate filter.
In higher dimensions (>3) the filter can be non continuous, which can complicate
the optimization. The kernel gradient is also scaled lower than other parameters,
which complicates the optimization.
How to compile the code
-----------------------
Dependencies:
* cmake http://www.cmake.org/
* Eigen (included)
* liblbfgs (included)
Linux, Mac OS X and Windows (cygwin):
mkdir build
cd build
cmake -D CMAKE_BUILD_TYPE=Release ..
make
cd ..
Windows
You're probably better off just copying all files into a Visual Studio
project
How to run the example
----------------------
An example on how to use the DenseCRF can be found in
examples/dense_inference.cpp. The example loads an image and some annotations.
It then uses a very simple classifier to compute a unary term based on those
annotations. A dense CRF with both color dependent and color independent terms
find the final accurate labeling.
Linux, Mac OS X and Windows (cygwin):
build/examples/dense_inference input_image.ppm annotations.ppm output.ppm
For example:
build/examples/dense_inference examples/im1.ppm examples/anno1.ppm output1.ppm
An example on how to unse the learning code can be found in
examples/dense_learning.cpp. The example loads a color image and ground truth
annotation. It then learn a CRF model with a logistic regression, a label comp
and Gaussian kernel.
Linux, Mac OS X and Windows (cygwin):
build/examples/dense_learning input_image.ppm annotations.ppm output.ppm
For example:
build/examples/dense_learning examples/im1.ppm examples/anno1.ppm output1.ppm
Please note that this implementation is slightly slower than the one used to
in our NIPS 2011 paper. Mainly because I tried to keep the code clean and easy
to understand.