new finalizer module

[xulman]

currently: CCA with choosing of the largests CC
try new: for candidates were large (>10%) collision were resolved, try to check all involved candidates as follows:

• find "octopus" by doing:

  • consider "neighbor matrix", N-by-N where N is the number of the involved candidates
  • fill position a,b by the number of pixels from 'a' that have pixel 'b' in their 8-neigh (3D) surrounding
  • octopus is a candidate whose line is "fat"
    • has more contacts to other candidates then the rest (that is, has more non-zero elems on its line)
    • non-zero items on its line are larger than what is usual for the rest

• cut (and remove) arms by doing:

  • erode by X px (width, X = 3px) all octopuses, eroded pixels are marked with special value
  • CCA on octopuses and keep largest CC for each
  • dilate back by X px all octopuses only over the special value'd pixels

This should remove "embracing around normal candidates by one that is oversegmenting a lot"...

[xulman]

Another tweak, motivated by a real (and rare) problem in some data, is that the current CCA is choosing the largest component but it should be choosing a component that is overlapping with the original TRA marker. Sometimes the same instance ID (in an instance segmentation result) is repeated, defining multiple-segments in this way, and not necessarily the largest component is the one that brought this multi-segments result for consideration w.r.t. to the currently examined TRA marker. Since this is the latest (cleaning) stage of the fusion, chosing a component that overlaps with TRA will also guarantee that the produced result overlaps with TRA (which again was sometimes seen in the fused data), see #7.