Daniel Dunbar, daniel@zuster.org

PDSample generates Poisson-disk sampling sets in the domain [-1,1]^2 using a variety of methods. See "A Spatial Data Structure for Fast Poisson-Disk Sample Generation", in Proc' of SIGGRAPH 2006 for more information.

The code should be portable to any platform with 32-bit float's and int's.

Windows: There is an included PDSample.sln for MSVS version 7. Unix: Type 'make' and hope for the best.

PDSample [-m] [-t] [-r ] [-M <multiplier=1>] [-N <minMaxThrows=1000>]

• -t

  Uses tiled (toroidal) domain for supporting samplers. The resulting point set will be suitable for tiling repeatedly in the x and y directions.

• -m

  Maximize the resulting point set. For samplers which do not already produce a maximal point set then this will use the Boundary sampling method to ensure the resulting point set is maximal.

• -r

  Apply the specified number of relaxations to the resulting point set. This requires that qvoronoi be in the path.

• -M

  For DartThrowing and BestCandidate methods this determines the factor to multiply the current number of points by to determine how many samples to take before exiting (DartThrowing) or accepting the best candidate (BestCandidate).

• -N

  This specifies a minimum number of samples that will be taken for the DartThrowing sampler. See below.

• DartThrowing

  Standard dart throwing. On each iteration the DartThrowing sampler will try min(N*multiplier,minMaxThrows) samples before termination. Note that for regular dart throwing where simply a maximum number of throws is used to determine the termination point, the multiplier should be set to 0.

• BestCandidate

  Mitchell's Best Candidate algorithm. Uses the multiplier argument.

• Boundary

  Dart throwing by maximizing boundaries.

• Pure

  Dart throwing using scalloped sectors.

• LinearPure

  Dart throwing using scalloped sectors but without sampling regions according to their probability of being hit.

• Penrose

  Ostromoukhov et al.'s sampling method using their quasisampler_prototype.h

• Uniform

  Random point generation. The number of samples to take is calculated as .75/radius^2 to approximately match the density of Poisson-disk sampling.

Point sets are output in a trivial binary format. The format is not intended for distribution and does not encode the endianness of the generating platform.

The format matches the pseudo-C struct below::

struct { int N; // number of points float t; // generation time float r; // radius used in generation int isTiled; // flag for if the set is tileable float points[N][2]; };

The format can be converted to JSON with the dump-points script.

Thanks to Ares Lagae for comments on a preliminary release of the code, Ostromoukhov et al. for making available their quasisampler implementation, as well as Takuji Nishimura and Makoto Matsumoto for their Mersenne Twister random number generator.