CPMC, version 1.0 Extension of the algorithm in the paper: Constrained Parametric Min-Cuts for Automatic Object Segmentation, Joao Carreira and Cristian Sminchisescu, CVPR 2010 -------------------- Overall idea The algorithm generates multiple energy minimization problems, by placing alternative square foreground seeds on a regular grid. It then solves all the problems for multiple parameters using parametric max-flow, and filters similar solutions. Finally it computes features on the remaining segments and ranks them. --------------------------- Differences to original paper: This version corresponds almost exactly to the one used on the VOC2010 challenge. The differences to the CVPR 2010 paper (and VOC2009 challenge) are: - uses a single soft background "frame"-seed, instead of 4 hard "frame" seeds. This reduces the number of parametric max flow (PMF) problems to solve to a fourth. - resize the image to 50%, by default, before solving the PMF problems. Also to improve speed. - can generate segments inside regions of interest defined by rectangles (uses by default a uniform grid of regions of interest, but supports using a detector ). - can generate seeds on a grid but with the precise locations defined by closeby superpixels. - ranking with linear regression on appearance and shape features, using random features to approximate the laplacian kernel, which is suitable for histogram features. See info in the paper "Random Features for Large-Scale Kernel Machines", Ali Rahimi. The approximation of the laplacian kernel was implemented by Fuxin Li and Catalin Ionescu. - Tuned to get high recall on VOC, so that it could potentially be used for detection. It's designed to rank from 800 to 100/200. The CVPR10/Pascal09 version was designed to rank down from 150 to 50/90. This package doesn't contain the code to learn the ranking models, but it includes two pretrained models, one on the training set and one on trainval set of the VOC2010 segmentation dataset. -------------------- Usage: Follow the cpmc_example.m to see how to use the code. See cpmc.m to see the options you can pass to generate segments. -------------------- External code included: - Parametric max flow implementations: http://riot.ieor.berkeley.edu/riot/Applications/Pseudoflow/parametric.html and http://research.microsoft.com/en-us/downloads/d3adb5f7-49ea-4170-abde-ea0206b25de2/default.aspx with custom additional mex files. - clipboxes.m from http://people.cs.uchicago.edu/~pff/latent-release4/ - pdist2.m: http://vision.ucsd.edu/~pdollar/toolbox/doc/index.html - functions from: http://www.csse.uwa.edu.au/~pk/Research/MatlabFns/ - Optimized chi-square code: http://mloss.org/software/view/64/ - An adaptation to segments of the pyramid hog code in http://www.robots.ox.ac.uk/~vgg/research/caltech/ - vlfeats: www.vlfeat.org/ - FH segmentation algorithm mex file from: from http://www.robots.ox.ac.uk/~ojw/software.htm for placing seeds inside the superpixels closer to locations in the regular grid. - Gauthier Fleutot immerge. - csift: http://staff.science.uva.nl/~ksande/research/colordescriptors/ - globalPb: http://www.eecs.berkeley.edu/~arbelaez/globalPb.html Not included but supported: - deformable part models: get in http://people.cs.uchicago.edu/~pff/latent/ and http://people.cs.uchicago.edu/~rbg/star-cascade/ You can define bounding boxes as regions of interest where additional seeds can be placed. See cpmc.m -------------------- Compatibility Tested on linux 64bits and 32bits, with Matlab 7.9 and 7.10. -------------------- Running Time and Requirements On a 2.00Ghz multicore cpu, running with 8 threads, it takes around 7 minutes to process an image. First 4 minutes are for getting the globalPB contour map and the last 3 minutes for the cpmc algorithm itself. Generating the segments takes around 1 minute the other 2 minutes are spent computing features and ranking the segments. Tested on a computer with 10Gb of RAM memory, without any problem. PB files for the examples in data/JPEGImages have been precomputed. -------------------- Support: Contact carreira@ins.uni-bonn.de or cristian.sminchisescu@ins.uni-bonn.de. Regarding bugs and comments on the code, contact preferentially the first. --------- Version history: 1.0 ------ - First public release. See the Changes file for all additions since the first release.