crf-motion-seg

Setup Dense CRF codebase

We use the py-dense-crf implementation inside a conda environment.

Inside the env:

pip install pydensecrf
pip install -U setuptools
pip install -U cython
pip install numpy matplotlib scipy scikit-image ipython jupyter

Apply CRF

The apply_crf.py file contains the Python code to apply dense CRF on motion segmentation datasets.

The run_*.sh shell scripts run the CRF segmentation on various datasets. The settings and paths should be modified inside these scripts before calling.

The apply_crf_image.py can be called from within a MATLAB script or from the terminal to do CRF segmentation for a single frame. It takes in one image file, one MAT file of raw segmentations and outputs the CRF segmentations in a MAT file at a specified location.

Setup with MATLAB calls on swarm2 cluster

The main purpose of apply_crf_image.py is to be able to call it from within a MATLAB loop -- with some file I/O overheads. The MATLAB instance needs to be run inside a conda environment. It can be a bit tricky to set this up on a SLURM cluster.

To setup on swarm2 (compute cluster using SLURM):

1. install miniconda (from your home directory on swarm2):

   wget https://repo.continuum.io/miniconda/Miniconda2-latest-Linux-x86_64.sh bash Miniconda2-latest-Linux-x86_64.sh

2. setup the conda environment:

    conda create --prefix ~/dense-crf-conda python=2.7
    source activate /home/arunirc/dense-crf-conda
    pip install pydensecrf

The last bit takes a long time ... without any output messages!

    conda install numpy
    conda install scipy
    conda install scikit-image

3. then check if this works on a single input:

   wget https://github.com/AruniRC/crf-motion-seg/archive/master.zip

   unzip master.zip

   cd crf-motion-seg-master/

   srun python apply_crf_image.py -i ./samples/bear01/bear01_0002.jpg -s ./samples/bear01/00002.mat -o ./samples/test_seg_2.mat -v

You should see ./samples/viz_raw_crf_bear01_0002.jpg -- the segmentation visualization image.

Basically, just run your SLURM scripts from inside the conda environment.