![]() |
Home | Libraries | People | FAQ | More |
#include <boost/math/interpolators/catmull_rom.hpp> namespace boost{ namespace math{ template<class Point> class catmull_rom { public: catmull_rom(const Point* const points, size_t num_points, bool closed = false, Real alpha = (Real) 1/ (Real) 2) catmull_rom(std::vector<Point>&& points, bool closed = false, Real alpha = (Real) 1/ (Real) 2) Real operator()(Real s) const; Real max_parameter() const; Real parameter_at_point(size_t i) const; Point prime(Real s) const; }; }}
Catmull-Rom splines are a family of interpolating curves which are commonly used in computer graphics and animation. Catmull-Rom splines enjoy the following properties:
The catmull_rom class provided
by boost creates a cubic Catmull-Rom spline from an array of points in any
dimension. Since there are numerous ways to represent a point in n-dimensional
space, the class attempts to be flexible by templating on the point type. The
requirements on the point type are discussing in more detail below, but roughly,
it must have a dereference operator defined (i.e., p[0]
is not a syntax error), it must be able to be dereferenced up to dimension -1, and p[i]
is of type Real, define value_type, and the free function size(). These
requirements are met by std::vector
and std::array. The basic usage is shown here:
std::vector<std::array<Real, 3>> points(4); points[0] = {0,0,0}; points[1] = {1,0,0}; points[2] = {0,1,0}; points[3] = {0,0,1}; catmull_rom<std::array<Real, 3>> cr(points.data(), points.size()); // Interpolate at s = 0.1: auto point = cr(0.1);
Memory conscious programmers may enjoy using the move constructor instead:
catmull_rom<std::array<Real, 3>> cr(std::move(points));
The spline can be either open or closed, closed meaning that there is some t such that P(t) = P(0). The default is open, but this can be easily changed:
// closed = true catmull_rom<std::array<Real, 3>> cr(points.data(), points.size(), true);
Inside catmull_rom, the Catmull-Rom
curve is represented as closed. This is because an open Catmull-Rom curve is
implicitly closed, but the closing point is
the zero vector. There is no reason to suppose that the zero vector is a better
closing point than the endpoint (or any other point, for that matter), so traditionally
Catmull-Rom splines leave the segment between the first and second point undefined,
as well as the segment between the second-to-last and last point. We find this
property of the traditional implementation of Catmull-Rom splines annoying
and confusing to the user. Hence internally, we close the curve so that the
first and last segments are defined. Of course, this causes the tangent
vectors to the first and last points to be bizarre. This is a "pick your
poison" design decision-either the curve cannot interpolate in its first
and last segments, or the tangents along the first and last segments are meaningless.
Since the routine internally represents the curve as closed, a question arises: Why does the user have to specify if the curve is open or closed? The answer is that the parameterization is chosen by the routine, so it is of interest to the user to understand the values where a meaningful result is returned.
Real max_s = cr.max_parameter();
If you attempt to interpolate for s
> max_s,
an exception is thrown. If the curve is closed, then cr(max_s)
= p0,
where p0 is the first point
on the curve. If the curve is open, then cr(max_s)
= pf,
where pf is the final point
on the curve.
The Catmull-Rom curve admits an infinite number of parameterizations. The default
parameterization of the catmull_rom
class is the so-called centripedal parameterization.
This parameterization has been shown to be the only parameterization that does
not form cusps or self-intersections within segments. However, for advanced
users, other parameterizations can be chosen using the alpha
parameter:
// alpha = 1 is the "chordal" parameterization. catmull_rom<std::array<double, 3>> cr(points.data(), points.size(), false, 1.0);
Finally, the tangent vector to any point of the curve can be computed via
double s = 0.1; Point tangent = cr.prime(s);
Since the magnitude of the tangent vector is dependent on the parameterization, it is not as meaningful as (say) arc-length parameterization. However, its direction is meaningful, so the user may wish to normalize this result.
The following performance numbers were generated for a call to the Catmull-Rom interpolation method. The number that follows the slash is the number of points passed to the interpolant.
Run on 2700 MHz CPU CPU Caches: L1 Data 32K (x2) L1 Instruction 32K (x2) L2 Unified 262K (x2) L3 Unified 3145K (x1) --------------------------------------------------------- Benchmark Time CPU --------------------------------------------------------- BM_CatmullRom<double>/4 20 ns 20 ns BM_CatmullRom<double>/8 21 ns 21 ns BM_CatmullRom<double>/16 23 ns 23 ns BM_CatmullRom<double>/32 24 ns 24 ns BM_CatmullRom<double>/64 27 ns 27 ns BM_CatmullRom<double>/128 27 ns 27 ns BM_CatmullRom<double>/256 30 ns 30 ns BM_CatmullRom<double>/512 32 ns 31 ns BM_CatmullRom<double>/1024 33 ns 33 ns BM_CatmullRom<double>/2048 34 ns 34 ns BM_CatmullRom<double>/4096 36 ns 36 ns BM_CatmullRom<double>/8192 38 ns 38 ns BM_CatmullRom<double>/16384 39 ns 39 ns BM_CatmullRom<double>/32768 40 ns 40 ns BM_CatmullRom<double>/65536 45 ns 44 ns BM_CatmullRom<double>/131072 46 ns 46 ns BM_CatmullRom<double>/262144 50 ns 50 ns BM_CatmullRom<double>/524288 53 ns 52 ns BM_CatmullRom<double>/1048576 58 ns 57 ns BM_CatmullRom<double>_BigO 2.97 lgN 2.97 lgN BM_CatmullRom<double>_RMS 19 % 19 %