/conservation_tracking_pipeline

Primary LanguageC++

C++ Pipeline to run Conservation Tracking

by Philipp Hanslovsky, Carsten Haubold, Martin Schiegg, David Stoeckel

University of Heidelberg, 2015

About

This pipeline runs the conservation tracking algorithm given a sequence of tiff images and pre-trained Random Forest classifiers. The classifiers can be extracted from previously trained ilastik projects for pixel classification and tracking, but every step will be rerun in the C++ pipeline (allows for batch-segmenting and tracking).

It was originally developed for the ISBI 2015 Cell Tracking Challenge, and is thus only operating on tiff sequences.

Compilation

Disclaimer: this was never compiled or tested on Windows!

The easiest way to get this compiled is by having a conda setup of ilastik-everything, let's assume that the respective conda environment is called ilastikdev. You also need the https://github.com/ilastik/ilastik-build-conda repository cloned somewhere locally. With your environment activated, install our gcc version as follows:

conda install -c ilastik gcc

Then you need to clone pgmlink yourself to check out a different (isbi_challenge_15) branch as follows

# in your favorite location:
git clone http://github.com/martinsch/pgmlink
cd pgmlink

# Just run the build script with the right environment setup
PREFIX=/path/to/your/miniconda/envs/ilastikdev \
PYTHON=/path/to/your/miniconda/envs/ilastikdev/bin/python2.7 \
CPU_COUNT=4 \
bash /path/to/your/ilastik-build-conda/pgmlink/build.sh

# now check out the proper pgmlink branch
git checkout isbi_challenge_15
cd build
make install

With this pgmlink version installed, you can clone, configure and build this C++ pipeline:

git clone https://github.com/chaubold/conservation_tracking_pipeline
cd conservation_tracking_pipeline
mkdir build
cd build
export LD_LIBRARY_PATH=/path/to/your/miniconda/envs/ilastikdev/lib # or DYLD_FALLBACK_LIBRARY_PATH on OSX
CC=/path/to/your/miniconda/envs/ilastikdev/bin/gcc \
CXX=/path/to/your/miniconda/envs/ilastikdev/bin/g++ \
cmake .. -DCMAKE_INSTALL_PREFIX=/path/to/your/miniconda/envs/ilastikdev -DCPLEX_ROOT_DIR=/path/to/your/cplex/root \
-DVIGRA_INCLUDE_DIR=/path/to/your/miniconda/envs/ilastikdev/include/ \ -DADDITIONAL_PGMLINK_INCLUDE_DIR=/path/to/your/pgmlink/build/include -DWITH_CHECKED_STL=OFF
make -j

Usage

Let's assume you trained a pixel classification project pc.ilp, exported the prediction maps, and then set up an automated tracking workflow tracking.ilp. We suggest to run tracking for only a small field of view to see that the chosen parameters yield sensible results. Save after running tracking to make sure the parameters are up to date in the file tracking.ilp

To create the configuration files for this C++ pipeline, use the following command:

python /path/to/this/repo/scripts/ilpToPipelineConfig.py --pixel-classification-project /path/to/pc.ilp --tracking-project /path/to/tracking.ilp --out /my/existing/out/dir

This will create a set of files in /my/existing/out/dir:

  • classifier.h5: the extracted random forests for pixel classification, object count detection and division detection
  • features.txt: the selected pixel classification features
  • object_count_features.txt: the selected features for object count classification
  • division_features.txt: the selected features for division classification
  • tracking_config.txt: a text file containing all other specified parameters

Using these files, the raw tiff images in /path/to/dataset/raw_images and two empty existing directories for the resulting segmentation (after pixel classification and thresholding) /path/to/dataset/segmentations, as well as resulting relabeled segmentations after tracking (where each tracked object is colored in its track ID) /path/to/dataset/result, the pipeline can be run as follows:

python /path/to/this/repo/build/isbi_pipeline /path/to/dataset/raw_images /path/to/dataset/segmentations /path/to/dataset/result /my/existing/out/dir/tracking_config.txt /my/existing/out/dir/classifier.h5 /my/existing/out/dir/features.txt /my/existing/out/dir/object_count_features.txt /my/existing/out/dir/division_features.txt

The file naming conventions for input and resulting files are subject to the Cell Tracking Challenge Guidelines

Notes:

  • The isbi_pipeline32 and tracking32 executables are using a label space of type uint32 whereas the default pipeline uses uint16. This is because one dataset in the ISBI challenge exceeded the number of labels representable by 16 bit.
  • The tracking executables are meant to work on previously obtained segmentations. Thus after running the isbi_pipeline once, you only need to run tracking again if you did not change of the parameters related to pixel classification.

References

  • M. Schiegg, P. Hanslovsky, B. X. Kausler, L. Hufnagel, F. A. Hamprecht. Conservation Tracking. Proceedings of the IEEE International Conference on Computer Vision (ICCV 2013), 2013.
  • Cell Tracking Challenge