Version 2.5:
        - We now include all of the LIBPMK extensions in this single package.
          This is because version numbers can get inconsistent, so it's better
          to keep them all in the same package. They still all compile
          independently.
        - Added libpmk_eth examples to the package.
        - Added the adaptive vocabulary tree extension (directory "avt")
        - Added a SURF binary to the feature extractor extension.
        - Added a demo UI to the ISM extension.
	- Fixed a bug in global-vg-pyramid-maker that caused inaccurate
          results due to not updating child nodes in the tree properly
          (thanks Popke Altenburg, Fabian Richter)
	- Fixed a bug in the ISM train-model example where it only used 1
          image as training data instead of all of them (thanks Guo Xi)
        - Fixed 64-bit compilation problems (thanks Gal Frishman)
        - Fixed compilation on Cygwin (thanks Arvid Terzibaschian, Tom Yeh)

Version 2.4:
	- Fix a logic bug in histogram.cc's distance computing functions. The
	  impact was that many distance computations were underestimated,
	  and, depending on the input, the output of these methods could have
	  been much less than what they should have been. (thanks Ruihua Ma)

Version 2.3:
	- Fix all compilation problems related to gcc 4.3.

Version 2.2:
	- UniformPyramidMaker can now use a vector of PointRefs in addition to
	  using a PointSet.
	- Read/WriteHeaderToStream now available for MultiResolutionHistogram.
	- Speedup in OnDiskPointSetList::GetPointRefs().
	- Histogram can now compute chi-squared distances and sum-of-square
	  distances.
	- SVMExperiment can now read/write itself to disk.
	- Adding RNNClusterer (reciprocal nearest neighbor clustering).

Version 2.1:
	- Minor speedup in MultiResolutionHistogram.
	- Default tools now use InputSpecificPyramidMaker instead of GlobalPyramidMaker.
	- Add a function to PointSetList to allow it to get PointRefs for just one image.
	- PointRefs can now be copied (copy constructor added)

Version 2.0 (breaking):
	- A lot of interface changes. Brief summary:
	  * "Feature" has been renamed to "Point"
	  * Simple getters and setters of a class are all lowercase.
	    For instance, if a class has a data member called size,
	    then the methods size() and set_size() will be there. If
	    the data member is a vector or list, the methods
	    member(int i), add_member(), and member_size() will be
	    present. Please see the documentation for details.
	  * The old Tree has been renamed to SparseTree (same with
	    TreeNode and SparseTreeNode). Taking the place
	    of Tree is a dense tree implementation.
	- The dense tree and associated TreeNode have been added, along
	  with PointTreeNode which is a TreeNode that has a Point in it.
	  A node in a dense tree no longer needs to store an entire path
	  to identify it; a node now has a single integer ID (rather than
	  LargeIndex which was used for sparse trees).
	- All tools have been updated to use the updated interface. Those
	  may be a good starting point.
	- HierarchicalClusterer now uses a dense tree representation and
	  is hopefully faster. Its new interface gives more intuitive
	  access to the tree structure.

Version 1.5:
	- Fix the behavior of GlobalVGPyramidMaker to explicitly compute
	furthest pairs for bin size calculations.(thanks Kristen Grauman)

Version 1.4:
	- KMeansClusterer will now return fewer than K clusters if the
	input data contains less than K unique points (in such a case it
	will return a number of clusters equal to the number of unique
	points).
	- Fix bug in Tree::DeleteNode() when deleting an only child.
	- Fix potential access error in Histogram::SortBins() on some
	platforms (thanks Yiqun Hu)
	- New default behavior of Feature's constructor is to set the
	initial weight to 1, not 0 (thanks Ravi Kiran)

Version 1.3:
	- Internal changes to Tree and TreeNode for faster processing.
	- Removed Tree::DepthFirstIterator and replaced it with
	PreOrderIterator and PostOrderIterator.
	- Added the ability to copy Trees.
	- Added Tree::DeleteNode(), which deletes a specified node as well
	as all children nodes.
	- Added a "merging" copy constructor to Tree which takes 2 trees
	and returns one big merged tree.
	- Minor fixes to Clusterer (more const member functions)
	- Clusterer has a new function, GetNumCenters().
	- Allow a UniformPyramidMaker to be constructed with a vector of
	PointRefs (in addition to a PointSetList)
	- PointSets now have the ability to serialize themselves (before,
	that was all handled by PointSetList).
	- The pointsets-to-uniform-pyramids tool takes less memory by
	keeping all of the PointSetLists on disk.
	- Added more user-friendly output to the tools.
	- Fixed some compilation errors on gcc-4.1.1.

Version 1.2:
	- Added depth-first and breadth-first iterators to Tree.
	- Added ReadFromFile() and WriteToFile() convenience functions for
	most classes that support serialization.

Version 1.1:
	- KernelMatrix now uses doubles to store data, rather than floats.
	- Added a Matrix class, which is just a rectangular 2D array of
	doubles.
	- Added a new constructor to Experiment, which allows one to pass
	in data in two separate pieces (training and testing), so you no
	longer have to compute kernel values between the test examples if
	you know which ones they are.
	- Added ProbSVMExperiment, another SVM experiment example, which
	reports the posterior probabilities of each class for each test
	example.
	- Added a new Tree class template and TreeNode, which lets you
	create custom tree-like structures. MultiResolutionHistogram has
	been rewritten as a wrapper around Tree, but its interface has not
	changed (all changes are transparent).

Version 1.0:
	- Initial release