The bare-bones code set for superpixel agglomeration and classifier training.
The methods implemeted in this code repo are described in the PloS ONE paper: Parag, T. et. al (2015). A Context-Aware Delayed Agglomeration Framework for Electron Microscopy Segmentation (http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0125825)
This repository borrows implementations of the graph and feature datastructures from Janelia FlyEM NeuroProof (https://github.com/janelia-flyem/NeuroProof), which I am also a contributor to.
Linux: Install miniconda on your workstation. Create and activate the conda environment using the following commands:
conda create -n my_conda_env -c flyem vigra=1.10 opencv
source activate my_conda_env
Then follow the usual procedure of building:
mkdir build
cd build
cmake -DCMAKE_PREFIX_PATH=[CONDA_ENV_PATH]/my_conda_env ..
The inputs and outputs for the following example are self explanatory. The option -threshold indicates the boundary predictor confidence at which the agglomeration should stop.
build/NeuroProof_stack -watershed watershed2_4class_600000_10_800_1000_1.0_2.h5 stack -prediction pixel_prediction2_4class_600000_10_800_1000_1.0_2.h5 stack -classifier int_classifier2_600000_1000_800_e5000_1.xml -output /result_600000_800_1_mthd0.35_e500_thd0.3_m4.h5 stack -algorithm 1 -threshold 0.3
The necesary files are uploaded to Dropbox instead due to filesize limitation.
https://www.dropbox.com/sh/y90ygc8nunpolyw/AADnypNkfF7om067Z3PaXJ9ca?dl=0
Note that the superpixel boundary classifier used in this example is not the optimal one. Therefore, the result may be suboptimal if one uses this particular classifier.
Contact: toufiq.parag@gmail.com