Search refactor

[carshadi]

Closes #53

(WIP)

Highlights (non-exhaustive):

• Faster, scalable path-finding and filling algorithms and data structures
• Encapsulate search cost functions and heuristics into objects, which are injected into general search routines. This enables mixing and matching different cost functions, heuristics, and search strategies in a decoupled, convenient way.
• Add a new probabilistic cost function for A* searches (taken from an A* implementation by the janelia workstation)
• Imagery is now represented and handled internally using imglib2 data structures, providing the ability to perform searches over arbitrary RandomAccessibleIntervals. This sets the stage for future migration to BigDataViewer for tracing through the GUI, improved integration with sciview (@kephale ) and already allows us to trace and fill over arbitrarily large image volumes via scripting (examples coming soon).
• 2 new, custom multi-scale vessel enhancement filters, implemented as ImageJ Ops. In my testing, these run faster than the Tubeness and Frangi vesselness filters currently available in ImageJ Ops. The new Frangi filter is implemented as close to the original paper as possible (well, attempted at least...), whereas the current Op Frangi filter is non-canonical in the way the hessian eigenvalues are combined. Additionally, the new Tubeness filter supports multi-scale integration, whereas the current Ops implementation is single-scale.
• Support for real-time local filtering of the image during A*-search, heavily reliant on imglib2-cache. This is intended as a generalization of the current hessian analysis functionality, which computes a single-scale Tubeness measure for voxels explored during searches. We can instead perform filtering using existing Ops, or define custom filters either as Ops or Consumers (approach heavily inspired by https://github.com/saalfeldlab/hot-knife/blob/master/src/main/java/org/janelia/saalfeldlab/hotknife/util/Lazy.java)
• Adaptive, local image statistics computations used by searches (optional), enables tracing over large volumes where computing the entire image statistics is prohibitive.
• Various UI improvements, including the new computation settings panel and improved Fill Manager interface.

Some Benchmarks

For these tests I used the following image as input (chunk of mouse cortex), with the test path shown in magenta (real length = 582.5 micrometers, # nodes = 1540) and the intensity histogram of the stack shown right. The start node is fixed for all iterations, and the path is traced between it and every subsequent node in the path, keeping track of the time-to-completion for the search and the count of explored nodes.

If we do this on the current master branch, we find that the searches quickly become slow as distance between the start and goal nodes increases (using TracerThread)

If we compare this with TracerThread and BiSearch (a new bidirectional A* implementation) after the proposed changes, we see at least a 30X speed improvement for both algorithms at the tail of the measurement while exploring similar numbers of nodes. Even though BiSearch explores more nodes on average than TracerThread, it is still significantly faster, with the gap between them growing as the search space becomes larger. This is due to some algorithm and data structure choices that reduce the computational burden compared to TracerThread, for example using only one image-like data structure to store the encountered nodes for the opposing searches instead of two. Unlike TracerThread, it is also guaranteed to produce the optimal path (assuming the implementation is correct), as it continues to refine the shortest-path graph even after the two searches initially meet (where the path may or may not be optimal).

Breaking Changes

We fundamentally changed how many of the tracing and Filling classes operate, which naturally led to API-break. We feel that this is necessary since many of the old method parameters are now obsolete, and the resulting API is simpler and more flexible. Users who have existing scripts which utilize tracing and filling classes (e.g., TracerThread, FillerThread) are encouraged to browse the new tracing sub-package of snt

There are almost certainly other items I'm forgetting...@tferr anything I missed?