This package computes the error between two segmentations of a volume. The metrics are either based on the Rand Index or Variation of Information, and features a few customizable options. These include:
-
Foreground Restriction- computing error over voxels where the second segmentation (presumed to be ground truth) != 0. This is applied as a default option.
-
Splitting '0' Segment into Singletons- The segment id of 0 often corresponds to a background segment, however it can also represent singleton voxels after processing by watershed. This option re-splits all segments labelled 0 into new segment ids, recovering the singletons. This is applied as a default option, and disabled by -no_split0
See the manual for detailed descriptions of each metric and its computation.
The main executable modules are:
score.py- One-shot scoring/comparison of two segmentationscurve.py- Computing error curves over the threshold a watershed MST (see Watershed)plot.py- Basic plotting functionality of error curves from curve.py
You can learn more about each of these modules by using a help flag.
python{3} score.py --help
python{3} curve.py --help
python{3} plot.py --help
metrics.py can also be used as an importable module for more flexible metric computation.
| Library |
|---|
| Cython >= 0.23.4 |
| python.tifffile |
| NumPy |
| Scipy |
| matplotlib |
make
NOTE: You will see a harmless warning when compiling the Cython functions. See (http://docs.cython.org/src/reference/compilation.html)
The codebase is now compatible with python3, in which case you can compile the Cython functions with
make python3
If you'd like to use segerror as a python module, rename/move the init.py file within the current directory after compilation