passessdetectionaccuracy

Synopsis

Description

The operator passessdetectionaccuracy calculates 2 errors to assess region detection accuracy of a segmentation result by comparison with a reference segmentation.

The two errors are:

• Recall error accounts for proportion of false negative regions.
• Precision error accounts for proportion of false positive regions.

Error values are between 0..1, where 0 means no error and 1 the worst result. A recall error of x means that x * 100 percent of the ground truth regions are not detected in the segmentation result. A precision error of x means that x * 100 percent of the segments do not detect any region. The result is stored in the output collection col_out which contains the two error values.

Discrepancy measures are computed using a matching graph between segments in the segmentation result and regions in the reference segmentation. Two types of graph matching algorithm can used depending on the parameter matching_algorithm_id: the first one allows for over and under-segmentation, and the second one is restricted to unique region-segment matching. In this graph, a segment S detects a region R if the overlapping area | R * S | is such that:

  | R * S |                                   | R * S |    
  --------- >= matching_threshold and  --------- >= matching_threshold
    | R |                                       | S |

Parameters

• matching_algorithm_id: the index of the matching algorithm to be used:
  • 0: for 1-n and n-1 matching. A segment in the segmentation result can merge several regions in the reference segmentation (under-segmentation), and a region in the reference segmentation can be split into several segments in the segmentation result (over-segmentation). However, a segment or a region cannot be both part of over-segmentation and under-segmentation.
  • 1: for 1-1 matching. A segment can only detect one region at the most, and a region can only be detected by one segment at the most.
• matching_threshold: minimum proportion of overlapping area between a region and a segment to accept a detection. The value is in the interval [0,1] where value x corresponds to a minimum of (x*100)% overlapping.

Inputs

• segmentation_result_in : a region map with segmentation result.
• reference_segmentation_in : a region map with reference segmentation.

Outputs

• col_out: a collection with the 2 error values.

Result

Returns SUCCESS or FAILURE.

Examples

Assesses detection accuracy of the segmentation result stored in 'tangram.pan' considering a minimum overlapping area of 50%:

passessdetectionaccuracy 0 0.5 resultimages/algo001/tangram.pan groundtruths/expert001/tangram.pan errors.pan
pvisu errors.pan

See also

Prototype C++