TemplRANSAC is a templated C++ generic RANSAC implementation using modern C++ template "duck-typing" instead of C++ abstract classes to specify the interface to be implemented by the user. The emphasis is on simplicity and readability in order to make RANSAC easier to understand.
The class is defined in a single header file Ransac.hh (in a namespace called templransac) with a single function:
template<typename Data, typename Model, typename Estimator>
double RANSAC(RANSACParams& parameters, const Estimator estimator, const Data& data, const size_t datasize,
const size_t sample_size, const size_t no_results, std::vector<std::pair<double, Model>>& results,
std::vector<std::vector<size_t>>& inliers, std::stringstream* errs = nullptr)The function returns a confidence value between 0 and 1.
The Data template parameter type acts as a data container container ie it contains containers with the actual data eg for PnP pose the data class could be defined as
struct PnPData
{
std::vector<cv::Point2f> &points2d;
std::vector<cv::Point3f> &points3d;
};The Model template parameter type contains the implementation specific information being optimised. The class should implement a copy constructor and an operator= (unless it is certain the trivial bitwise copy will do the right thing).
The Estimator template parameter type contains the implementation specific sample estimation and error calculations. It should implement two methods namely:
const int estimate(const D &samples, const std::vector<size_t> &sampleIndices,
RANSACParams ¶meters, std::vector<M> &fitted_models) const { return 0; }
const void error(const D& samples, const std::vector<size_t>& sampleIndices,
M& model, std::vector<double>& errors,
std::vector<size_t>& inlier_indices, std::vector<size_t>& outlier_indices,
double error_threshold) const {};The methods are called by template 'duck-typing' so the class does not need to be derived from a base class. In both methods the sampleIndices vector contains indices into the samples data containers which indicate the data items which the methods should use when executing their calculations. The estimate method returns the number of models generated (usually 1 but many pose algorithms solving quadratic or higher polynomials may return more than 1.) (The RansacHelp.h header does define a abstract base class for the estimator for use during development to make it easier to check the method signatures when using an IDE but it should be removed or commented out in release builds.)
parameters - The RANSAC parameters. In the interests of readability and ease of use the number of RANSAC parameters are held to a minimum and reflect those described in 4.7 of Multiple View Geometry in Computer Vision (2nd Ed):
- error_threshold: Error threshold used to distinguish between inliers and outliers. As the user defines the methods using this parameter it is up to the implementation to decide if it is a squared error or not (when an L2 norm is used).
- sample_inlier_probability: The probability that at least one sample is free of outliers
- inlier_probability: The probability that a given data item is an inlier
estimator The RANSAC estimator as described in the Estimator template type above. data The data container container as described in the Data template type above. datasize The size of the data in the Data container sample_size The minimum sample size to be used by RANSAC no_results The number of results to return (allows multiple solutions to be returned sorted by the cost. results The RANSAC results as a vector of pairs of cost and the best Model. If no_results > 1 then the results are sorted in descending order. In both 1 and > 1 cases the best result in results[0]. inliers The inliers as a vector of vectors corresponding to the models for the same index in results. errs A pointer to a stringstream which, if not null, will contain a human readable form of any errors.
The src/samples directory contains the standard line fitting RANSAC example as well as an OpenCV PnP example.