This repository contains part of the code that was used for the experiments in
C.Dann, P. Gehler, S. Roth, S. Nowozin Pottics -- The Potts Topic Model for Semantic Image Segmentation DAGM 2012
We based the implementation of our training scheme on the SVM^struct with latent variables of C.-N. Yu and T. Joachims. The source code may not be distributed without the permission of the authors. Therefore, we provide a diff located in latentsvm/latentsvm.patch that contains the changes to the Latent SSVM implementation template available at http://www.cs.cornell.edu/~cnyu/latentssvm/ . To obtain our training code, download this template, extract it and apply our patch file with the patch command (on Unix systems) by executing the following command in the extracted directory
patch -p1 -i path/to/pottics/repo/latentsvm/latentsvm.patch
Please adapt the path to your copy of this repository.
Afterwards recompile the package. For instructions, see
the README contained in the Latent SSVM template.
The input files for SVM training can be generated with the Python function
util.binfile.export_svmstruct_dataset.
To generate the proposal segmentations, the original CPMC implementation was used. We thank again the authors of CPMC, who made their code publicly available. You can download it from http://sminchisescu.ins.uni-bonn.de/code/cpmc/ . It can be used to store the best binary segmentation masks in a mat file. See the Data section for details, where to store the masks.
We compute the unary factor connected to each region by accumulating the predictions for each pixel in that region. We use precomputed predictions for the VOC2010 dataset, that are available at http://graphics.stanford.edu/projects/densecrf/unary/ ("compressed probabilities"). Thanks again to Philipp Krähenbühl and Vladlen Koltun for providing them. The predictions have to be decompressed as explained on the website. For information where to put the decompressed pixel predictions, see the Data format section.
The region generation and prediction of segmentations for images was implemented in Python
with Numpy. See the script generate_paper_plots.py for examples how to use the code, i.e.
make predictions for the baselines or Pottics.
The implementation makes heavy use of harddisk caching. This is done either explicitely by
storing results to mat files or implicitly by the joblib library.
For example, regions for an image are created from the segments automatically,
when you access the regions-property of an image the first time and is then stored to disk.
Whenever you access the property again, it is loaded from disk instead of recomputed:
import dataset as ds
# compute the regions for the first image of the VOC2010 dataset
# this may take a while
reg1 = ds.voc2010[0].regions
# access the regions again. This is fast.
reg2 = ds.voc2010[0].regions
assert(np.all(reg1 == reg2))
A set of images that is used for training or prediction is called dataset. All data associated
with it are stored wihin a single directory. The location of this folder is specified in the
module dataset as a member of each dataset subclass (e.g.:
dataset.VOC2011Dataset.path = ../../datasets/formatted/voc2011). You can adopt this path easily to
your local circumstances. We will refer to this path as {dspath}. Each image is identified by a
unique name ({name}) within a dataset.
The following filesystem structure is assumed:
{dspath}/images/{name}.jpg -- the image itself {dspath}/ground_truth/{name}.png -- the ground truth segmentation (if available) {dspath}/densecrf/{name}.unary -- the decompressed unary predictions {dspath}/cpmc/MySegmentsMat/{name}/top100_masks.mat -- the binary segmentation masks generated by CPMC {dspath}/cpmc/MySegmentsMat/{name}/scores.mat -- the ranking scores for each CPMC segmentation