An implementation for an efficient segmentation with a supermodular loss function. This implementation also allows user-defined loss functions and user-defined optimisations for the relative loss maximization and inference procedure.
For more details: Yu, J. and M. B. Blaschko: Efficient Learning for Discriminative Segmentation with Supermodular Losses. BMVC, 2016.
- Usage
(1) Dataset:
This implementation uses the dataset provided by:
V. Gulshan, C. Rother, A. Criminisi, A. Blake and A. Zisserman Geodesic Star Convexity for Interactive Image Segmentation, CVPR, 2010.
Run directly main.m first. It by defaut runs on a preprocessed and downsampled dataset "Sampled690.mat".
One can change the dataset in mainInit.m in order to use the OriginalData.mat, which takes longer time.
(2) main functions:
- main.m: overall running function;
- mainInit.m: parameter settings, load/prepare the data, splitting train/val/test, etc.;
- mainTrain.m: trainning cases, save trained weight vector;
- implement_SOSVM_Learning.m: SVM settings, callback functions;
- GeneralCalassifier_SOSVM.m: SVM oracle, cutting plane method;
(3) customer loss functions and its optimization:
(i) Define your loss function in customLossFunction.m following the specific format.
(ii) Use your loss function in mainInit.m with the name your defined.
(iii) '8connected' is the loss in our paper.
- Rerun the prepocessing
We followed the feature extraction procedure as in:
Anton Osokin and Pushmeet Kohli. Perceptually inspired layout-aware losses for image segmentation. In ECCV, 2014.
Download the dataset:
- Images: http://www.robots.ox.ac.uk/~vgg/data/iseg/data/images.tgz
- Ground truth: http://www.robots.ox.ac.uk/~vgg/data/iseg/data/images-gt.tgz
- Brush strokes: http://www.robots.ox.ac.uk/~vgg/data/iseg/data/images-labels.tgz
(i) Add the dataset at "./generateUnary/Dataset/";
(ii) run ./generateUnary/main.m and save the output X and Y.
Note: this processing can take very long time.
- Reference
This code uses several packages. Please consider citing :
(1) SFO toolbox Andreas Krause. SFO: A toolbox for submodular function optimization. JMLR, 11:1141–1144, 2010.
(2) GCMex & BKMatlab - MATLAB wrapper for the Boykov-Kolmogorov graph cuts
Yuri Boykov and Vladimir Kolmogorov. An experimental comparison of min-cut/maxflow algorithms for energy minimization in vision. T-PAMI, 26(9):1124–1137, 2004.
Yuri Boykov, Olga Veksler, and Ramin Zabih. Efficient approximate energy minimization via graph cuts. T-PAMI, 20(12):1222–1239, 2001.
B. Fulkerson, A. Vedaldi, and S. Soatto. Class segmentation and object localization with superpixel neighborhoods. In ICCV, 2009.
(3) Geodesic star convexity
V. Gulshan, C. Rother, A. Criminisi, A. Blake and A. Zisserman, Geodesic star convexity for interactive image segmentation. In Proceedings of Conference on Vision and Pattern Recognition (CVPR 2010).
Any questions please contact: jiaqian[dot]yu[at]centralesupelec[dot]fr
-- Jiaqian Yu @ 2016