Documentation

bboxPrecisionRecall

Compute bounding box precision and recall against ground truth

Description

example

[precision,recall] = bboxPrecisionRecall(bboxes,groundTruthBboxes) measures the accuracy of bounding box overlap between bboxes and groundTruthBboxes. Precision is a ratio of true positive instances to all positive instances of objects in the detector, based on the ground truth. Recall is a ratio of true positive instances to the sum of true positives and false negatives in the detector, based on the ground truth.

If the bounding box is associated with a class label, precision and recall contain metrics for each class. If the bounding box is also associated with a confidence score for ranking, use the evaluateDetectionPrecision function.

[precision,recall] = bboxPrecisionRecall(bboxes,groundTruthBboxes,threshold) specifies the overlap threshold for assigning a given box to a ground truth box.

Examples

collapse all

Create two ground truth boxes.

groundTruthBoxes = [2 2 10 20; 80 80 30 40];

Create three boxes for evaluation.

boundingBoxes = [4 4 10 20; 50 50 30 10; 90 90 40 50];

Plot the boxes.

figure
hold on
for i=1:2
rectangle('Position',groundTruthBoxes(i,:),'EdgeColor','r');
end
for i=1:3
rectangle('Position',boundingBoxes(i,:),'EdgeColor','b');
end Evaluate the overlap accuracy against the ground truth data.

[precision,recall] = bboxPrecisionRecall(boundingBoxes,groundTruthBoxes)
precision = 0.3333
recall = 0.5000

Define class names.

classNames = ["A","B","C"];

Create bounding boxes for evaluation.

predictedLabels = {...
categorical("A",classNames); ...
categorical(["C";"B"],classNames)};
bboxes = {...
[10 10 20 30]; ...
[60 18 20 10; 120 120 5 10]};
boundingBoxes = table(bboxes,predictedLabels,'VariableNames',...
{'PredictedBoxes','PredictedLabels'});

Create ground truth boxes.

A = {[10 10 20 28]; []};
B = {[]; [118 120 5 10]};
C = {[]; [59 19 20 10]};
groundTruthData = table(A,B,C);

Evaluate overlap accuracy against ground truth data.

[precision,recall] = bboxPrecisionRecall(boundingBoxes,groundTruthData)
precision = 3×1

1
0
1

recall = 3×1

1
0
1

Input Arguments

collapse all

Bounding boxes, specified as one of the following. M is the number of bounding boxes.

• For single-class bounding boxes, bboxes can be an M-by-4 matrix, or a table with M rows and one column. Each row of the matrix or element in the table represents a bounding box, specified in the format [x y width height], where x and y correspond to the upper left corner of the bounding box.

• For multi-class bounding boxes, bboxes is a table with M rows and two columns. Each element in the first column represents a bounding box, specified in the format [x y width height]. The second column contains the predicted label for each box. The label must be a categorical type defined by the variable (column) names of the groundTruthBboxes table.

Data Types: single | double | int8 | int16 | int32 | uint8 | uint16 | uint32

Ground truth bounding boxes, specified as one of the following. M is the number of ground truth bounding boxes.

• For single-class bounding boxes, groundTruthBboxes can be an M-by-4 matrix, or a table with M rows and one column. Each row of the matrix or element in the table represents a bounding box, specified in the format [x y width height], where x and y correspond to the upper left corner of the bounding box.

• For multi-class bounding boxes, groundTruthBboxes is a table with M rows and multiple columns. Each column represents a different class, and the column name specifies the class label. Each element in the table has the format [x y width height].

Data Types: single | double | int8 | int16 | int32 | uint8 | uint16 | uint32

Overlap threshold for assigned a detection to a ground truth box, specified as a numeric scalar. The overlap ratio is computed as the intersection over union.

Output Arguments

collapse all

Precision values from each detection, returned as a numeric scalar for single-class bounding boxes. For multi-class bounding boxes, precision is returned as a numeric vector consisting of metrics for each class. The class order follows the same column order as the groundTruthBboxes table.

Recall values from each detection, returned as a numeric scalar for single-class bounding boxes. For multi-class bounding boxes, recall is returned as a numeric vector consisting of metrics for each class. The class order follows the same column order as the groundTruthBboxes table.