This is an archive of the old MediaWiki-based ImageJ wiki. The current website can be found at imagej.net.

ImgLib2 Algorithms

Conventions for algorithm development in ImgLib2

  • Algorithms should either implement java.lang.Runnable if they do not return an object or java.util.concurrent.Callable if they do; no further interface hierarchy will provided for now
  • Multithreading will be solved through the ops package for easily parallelizable algorithms
  • More complicated algorithms that can provide significantly increased performance by implementing their own multithreading scenarios should implement the MultiThreading interface which allows to adjust the number of threads
  • Algorithms should take RandomAccessible, RandomAccessibleInterval and/or Iterable as input, Img should be only used as temporary data structure
    • RandomAccessible should be used if the OutOfBounds is defined externally (e.g. for gaussian convolution), an Interval should be passed as an extra parameter which defines the area that should be processed
    • RandomAccessibleInterval should be used if the algorithm defines its own OutOfBounds strategy (e.g. FFT) outside the given boundary; an Interval should be passed as an extra parameter which defines the area that should be processed
    • Iterable should be used if per-pixel operations are performed, here multithreading is delegated upstream
  • JUnit tests should be placed in the respective subpackages and must not require any additional imports
  • Human tests that might display images, show graphs or require any other imports than the algorithm itself should be placed in the ImgLib2-Tests project following the path convention of the algorithm
  • Standard images for Human tests will not be part of the git repository but will be available as http-link; a special opener class net.imglib2.io.ImgIOUtils will provide download and permanent caching in the temporary directory of each computer
  • Algorithms will have no direct dependency on the ImageJ2 plugin framework as they are generic; extra classes will provide access to the functionality of those algorithms
  • A Benchmarker class will be implemented to measure the performance for different containers, types, operating systems and machines