ngs333 / cmt Goto Github PK

There are several compile time macro definitions hard-coded with the the #define statement. These should be
be made compile time configuration options.

The metric trees should provide the options of loose (or non-tight) bounding intervals. This is particularly so if bounding distances of metrics calculated as doubles are stored in 32 bit floats. Without such a capability, the radius used in range search needs to have an extra epsilon (small quantity) increase of value, and in knn search is is not as clear. The loose bounding intervals should somehow be made only as loose as possible to not miss any results while not providing the extra epsilon with the query range. Possible, the C/C++ nextafter functions can be used for this.

Tree nodes should store the internal boundary interval sizes. Although this does not affect the correctness of the result set, it effects the query efficiency. During the process of tree construction tree nodes are storing only the distance to the nearest and the farthest objects (of child nodes) in the common distance interval. However, the distance to the furthest in the left child, and the distance to the nearest in the right child, are not being stored. They are available as a product of partitioning, and are very useful in discriminating between left and right child info for the purposes of pruning. Preliminary tests suggest that the information greatly influences the performance of the common metric tree at smaller query radius, while slightly influencing the performance of the CMT.