A simple C++ implementation of the mean shift clustering algorithm with the following properties and restrictions:
- Comprised of a single header file for the main algorithm, with additional headers for extra utilities.
- Can use OpenMP to parallelize the execution.
- Templated to support different scalar types.
- Supports custom kernel, distance and bandwidth estimator functions.
- Template specialization system to support any container/structure type.
- Requires C++11 support.
The most important function is mean_shift_cluster. The following is a simple example illustrating its usage:
List<Point<Scalar>> points = get_points();
std::vector<msc::Cluster<Scalar>> clusters = msc::mean_shift_cluster<Scalar>(
std::begin(points), std::end(points), 3, metric, kernel, estimator);The main algorithm is in msh.h. Implementations of some common metrics and kernels and a test estimator is in msc.metrics.h, msc.kernels.h and msc.estimators.h, respectively, but any custom functor or function that implements the appropriate signature is valid. The signatures are:
double metric(const Scalar* a, const Scalar* b, int dim);
double kernel(double distance);
double estimator(const Scalar* point,
InputIterator first, InputIterator last, int dim, Metric metric);To account for custom containers for point data, an msc::Accessor functor can be specialized to the desired type and the compiler will pick the correct specialization to access the container values. The signature is:
struct Accessor<Scalar, Point>
{
static const Scalar* data(const Point& point);
}Specializations for common types are included in msc.accessors.h, so that things like raw arrays, std::arrays and std::vectors work just by including this header (Note: it must be included after the main msc.h header for the compiler to know about the base template beforehand).
A helper header msc can be used to include all these headers in a single line.
A generic calculator is included in the file main.cpp that reads points from a file or the standard input and dumps the clustered points to the standard output. Some tests are included in the following files:
test_custom_struct: Exemplifies the use of a custom structure (Point3) to store points.test_1d_flat_vector: Uses a flat vector to store 1D points. This configuration works thanks to one of the accessors included inmsc.accessors.h.
The script test.sh uses the main executable to read the dataset in test.txt (obtained from here) and plots the results with gnuplot.