find_uncalib_descriptor_model — Find the best matches of a descriptor model in an image.
The operator find_uncalib_descriptor_model finds the best matches of a descriptor model ModelID in Image. The descriptor model must have been created previously by calling create_uncalib_descriptor_model, create_calib_descriptor_model or read_descriptor_model.
A match is only accepted if its score exceeds the value of MinScore. This criterion is based on the 'inlier_ratio' score that is described in details below. The main result of the operator find_uncalib_descriptor_model for each match is a 3x3 matrix HomMat2D, which describes a 2D projection of model points to search image points and is represented by a tuple of 9 elements in row-major order. If more matches of the searched object (template) appear and pass the MinScore criterion, the resulting multiple homographies are concatenated. The number of objects actually found is then equal to NumObjects = |HomMat2D|/9.
The detection process is divided into three parts. First, interest points are extracted from the search image (only inside the domain of the search image). This is done using the point detector and its parameters selected once during the generation of the model. However, DetectorParamName and DetectorParamValue can be used to specify different detector parameter values during the find_uncalib_descriptor_model call. By changing these parameters, it is possible to adjust to illumination changes between the model generation and the online detection. However, it is recommended to use the same values used to create the model (pass an empty tuple).
The second step of the detection process is to calculate correspondences between the model points and the points that were detected. The run time parameters of the descriptor can be adjusted with DescriptorParamName and DescriptorParamValue:
is the minimal classifier score for an interest point to be regarded as a potential match. The score function is between 0.0 and 1.0, but typically only values between 0.0 and 0.1 make sense. Increasing 'min_score_descr' can increase significantly the detection speed. Note, however, that using 'min_score_descr' might have negative effect on the robustness of the detection process, especially when only few points can be found. Typical values are [0.0 .. 0.1], default value is 0.0.
enhances the accuracy of the object recognition when switched on. Note that it increases the computational costs up to 10% in some cases. Possible values are ['on', 'off'], default value is 'on'.
The last step is the estimation of a homography that describes the point correspondences. The homography is a 2D projection, which describes a transformation from model points to points in Image. Here, Natural 3D Markers (N3Ms) are utilized to identify robustly the point corresondences (see references).
Additionally to the estimated homography in HomMat2D, the operator returns one or more Score estimations per object instance as specified by the user in a tuple ScoreType. Currently the following values for ScoreType are supported:
number of point correspondences per instance. An object instance should be considered good, if it has 10 or more point correspondences with the model. Fewer points are insufficient, because any random 4 point correspondences define a mathematically correct homography between two images.
the ratio of the number of point correspondences to the number of model points. Although this value can have values of [0.0 .. 1.0], it is rather unlikely that this ratio can reach 1.0. Yet, objects having inlier ratio less than 0.1, should be disregarded.
Note that the resulting scores for more than one object instance will be concatenated in Score, such that |Score| = NumObjects*|ScoreType|.
The point correspondences for each object can be queried with get_descriptor_model_points.
Note that the domain of the search image should contain the whole object to be searched for because interest points are only extracted inside the domain of the search image. This means that if the domain does not contain the full object to be searched for, the resulting Score will decrease. Note also that matches may be found even if the reference point (origin) of the model lies outside of the domain of the search image. Both is in contrast to shape-based matching, where the domain of the search image defines the search space for the reference point of the model.
Input image where the model should be found.
The handle to the descriptor model.
The detector's parameter names.
Default value: 
List of values: 'alpha', 'check_neighbor', 'mask_size_grd', 'mask_size_smooth', 'min_check_neighbor_diff', 'min_score', 'radius', 'sigma_grad', 'sigma_smooth', 'subpix', 'threshold'
Values of the detector's parameters.
Default value: 
Suggested values: 0.08, 1, 1.2, 3, 15, 30, 1000, 'on', 'off'
The descriptor's parameter names.
Default value: 
List of values: 'guided_matching', 'min_score_descr'
Values of the descriptor's parameters.
Default value: 
Suggested values: 0.0, 0.001, 0.005, 0.01, 'on', 'off'
Minimum score of the instances of the models to be found.
Default value: 0.2
Suggested values: 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0
Typical range of values: 0 ≤ MinScore ≤ 1
Maximal number of found instances.
Default value: 1
Suggested values: 1, 2, 3, 4
Restriction: NumMatches >= 1
Score type to be evaluated in Score.
Default value: 'num_points'
List of values: 'inlier_ratio', 'num_points'
Homography between model and found instance.
Score of the found instances according to the ScoreType input.
create_uncalib_descriptor_model (ImageReduced,'harris',,, \ ,,42,ModelID) get_descriptor_model_params (ModelID,DetectorType, \ DetectorParamName,DetectorParamValue, \ DescriptorParamName,DescriptorParamValue) write_descriptor_model (ModelID,'simple_example.dsm') clear_descriptor_model (ModelID) read_descriptor_model ('simple_example.dsm',ModelID) find_uncalib_descriptor_model (SearchImage,ModelID,,,,,0.2,1, \ ['num_points','inlier_ratio'],HomMat2D,Score) clear_descriptor_model (ModelID)
create_uncalib_descriptor_model, create_calib_descriptor_model, read_descriptor_model
create_uncalib_descriptor_model, create_calib_descriptor_model, find_calib_descriptor_model, get_descriptor_model_points
S. Hinterstoisser, S. Benhimane, and N. Navab: “N3M: Natural 3D Markers for Real-Time Object Detection and Pose Estimation.“ IEEE 11th International Conference on Computer Vision, 2007. pp. 1-7, ICCV 2007.