Supplementary Material + TextonBoost code
If you're using any of the material or code on this page, cite our paper Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials .
Here is an implementation of textonboost that should give you about 83-84% on MSRC.
The code is provided as is, so please don't write me about getting it to run properly.
I haven't tested it on any OS but linux.
It's under BSD license.
TextonBoost is probably patented.
In any case it should be used for research only.
How to use:
This will create a bunch of binary files that can be loaded as "Image<float>". The values are either energies or probabilities depending on the RAW_BOOSTING_OUTPUT flag (on: energies, off: probabilities).
- Mess with settings.h. Give it the path to the MSRC or VOC dataset.
- Build it using CMake (dependencies: QT, Eigen3, TBB) [out of place building is encouraged]
- Compute the textons : ./textonize_all.sh
- Train the classifier: build/src/textonboost data/msrc.dat data/msrc_color.dat data/msrc_filterbank.dat data/msrc_hog_l.dat data/msrc_location.dat
- Evaluate the classifier: build/src/evaluate data/msrc.dat data/msrc_color.dat data/msrc_filterbank.dat data/msrc_hog_l.dat data/msrc_location.dat path/to/a/large/harddrive
Additional material and dataset:
We split the MSRC training data in the following way. The accurate annotations can be found here.
To get textonboost to yield a decent result on VOC 2010 you need to train a bounding box classifier and use its response as a texton for texton boost.
We used Discriminatively Trained Deformable Part Models.