imageMetrics

Evaluate per-image object detection performance metrics

Since R2026a

collapse all in page

Syntax

imageSummary = imageMetrics(metrics)
[imageSummary,imagePredictions,imageGroundTruth] = imageMetrics(metrics)
[___] = imageMetrics(metrics,imageIndices)
[___] = imageMetrics(___,Name=Value)

Description

imageSummary = imageMetrics(metrics) evaluates the object detection metrics, metrics, for each image in the data set.

example

[imageSummary,imagePredictions,imageGroundTruth] = imageMetrics(metrics) evaluates the object detection metrics, metrics, for each image in the data set, and returns the performance summary, predictions, and ground truth information for each image.

example

[___] = imageMetrics(metrics,imageIndices) evaluates the object detection metrics, metrics, for images in the data set specified using the image indices, imageIndices.

example

[___] = imageMetrics(___,Name=Value) specifies options that configure per-image metrics using name-value arguments, in addition to the input arguments from the previous syntaxes. For example, ClassNames=["person" "car"] specifies to evaluate per-image metrics for the person and car classes.

example

Examples

collapse all

Open Live Script

Load Detection Results and Ground Truth Annotations

Load precomputed detection results and ground truth labels. The ground truth data is a subset of the Indoor Object Detection Dataset created by Bishwo Adhikari, and is in the form of a boxLabelDatastore that returns the object bounding boxes and their corresponding class labels for every image [1]. The MAT files containing the precomputed results and annotations are attached to this example as supporting files. To train an object detector on a custom dataset and generate the detection results, see Multiclass Object Detection Using YOLO v2 Deep Learning.

load("yolov2IndoorObjectDetectorAnnotation100.mat");
load("yolov2IndoorObjectDetectorResults100.mat");

Evaluate and Summarize Object Detection Metrics

Evaluate the object detection results against the ground truth using the evaluateObjectDetection function. Compute the metrics summary across the dataset and classes using the summarize function.

evalMetrics = evaluateObjectDetection(results,blds,[0.5,0.75,0.9]);
[summaryDataset,summaryClass] = summarize(evalMetrics);
disp(summaryDataset)
    NumObjects    mAPOverlapAvg    mAP0.5     mAP0.75     mAP0.9 
    __________    _____________    _______    _______    ________

       171           0.40012       0.74165    0.44671    0.012002

disp(summaryClass)
                        NumObjects    APOverlapAvg     AP0.5     AP0.75       AP0.9   
                        __________    ____________    _______    _______    __________

    exit                    53          0.60733        0.9619    0.82106      0.039027
    fireextinguisher        49          0.54426       0.93878    0.68903     0.0049712
    chair                   55          0.35992       0.88449    0.19463    0.00062696
    clock                   13           0.4891       0.92308    0.52885      0.015385
    trashbin                 1                0             0          0             0
    screen                   0              NaN           NaN        NaN           NaN
    printer                  0              NaN           NaN        NaN           NaN

The summary across classes shows that the "trashbin" class has only one ground truth object in this evaluation dataset, while the "screen" and "printer" classes do not have any ground truth objects present.

Evaluate Per-Image Metrics

Compute per-image metrics for the data set.

imageSummary = imageMetrics(evalMetrics);

Display the first ten images and their summarized per-image metrics.

disp(imageSummary(1:10,:) )
    ImageIndex    NumPredictedObjects    NumGroundTruthObjects    TP    FP    FN    Precision    Recall 
    __________    ___________________    _____________________    __    __    __    _________    _______

         1                 2                       2              2     0     0            1           1
         2                 2                       2              2     0     0            1           1
         3                 1                       1              1     0     0            1           1
         4                 0                       2              0     0     2          NaN           0
         5                 3                       2              2     1     0      0.66667           1
         6                 2                       3              2     0     1            1     0.66667
         7                 2                       2              2     0     0            1           1
         8                 2                       2              2     0     0            1           1
         9                 2                       2              2     0     0            1           1
        10                 3                       2              2     1     0      0.66667           1

Plot the histogram of false positive (FP) and false negative (FN) counts for each image.

figure;
subplot(1,2,1)
histogram(imageSummary.FP)
xlabel("False Positives Per Image")
ylabel("Frequency")
grid on
subplot(1,2,2)
histogram(imageSummary.FN)
xlabel("False Negatives Per Image")
ylabel("Frequency")
grid on

Most images have very few mistakes. The FP histogram shows that 8 images have one false positive detection. The FN histogram shows that 6 images have one false negative, and only 3 images have two false negatives.

[1] Adhikari, Bishwo; Peltomaki, Jukka; Huttunen, Heikki. (2019). Indoor Object Detection Dataset [Data set]. 7th European Workshop on Visual Information Processing 2018 (EUVIP), Tampere, Finland.

Open Live Script

Load precomputed detection results and ground truth labels. The ground truth data is a subset of the Indoor Object Detection Dataset created by Bishwo Adhikari [1]. The MAT files containing the precomputed results and annotations are attached to this example as supporting files.

load("yolov2IndoorObjectDetectorAnnotation100.mat");
load("yolov2IndoorObjectDetectorResults100.mat");

Evaluate Object Detection Results

Evaluate the object detection results against the ground truth using the evaluateObjectDetection function. Specify three overlap thresholds for the evaluation: 0.5, 0.75, and 0.9.

evalMetrics = evaluateObjectDetection(results,blds,[0.5 0.75 0.9]);

Analyze Detector Performance Across Object Size and Class

Compute per-image metrics for the data set.

[imageSummary,imagePredictions,imageGroundTruths] = imageMetrics(evalMetrics);

To analyze object size- and class-based detector performance, create a visual representation of the number of true positives, false positives, false negatives for the entire data set. First, concatenate all prediction tables from all images into a single table tblPreds.

tblPreds = vertcat(imagePredictions{:});

Create a logical array indicating which predictions are matched to ground truth objects, or are true positives (TPs).

isMatchedPred = logical(tblPreds.IsMatched);

Extract true positives (TPs) by keeping only the Label (class name) and Area (object size) columns for rows where predictions are matched to ground truth.

tp = tblPreds(isMatchedPred,["Label" "Area"]);

Extract false positives (FPs) by keeping only the Label (class name) and Area (object size) columns for rows where predictions are not matched to ground truth.

fp = tblPreds(~isMatchedPred,["Label" "Area"]);

Concatenate all ground truth tables into tblTruths, and create a logical array indicating which ground truths were matched by predictions. Extract false negatives (FNs) by keeping only Label and Area for ground truths not matched to any prediction.

tblTruths = vertcat(imageGroundTruths{:});
isMatchedGT = logical(tblTruths.IsMatched);
fn = tblTruths(~isMatchedGT,["Label" "Area"]);

Plot TPs as black circles using a swarm chart, with class name on the x-axis and object size on the y-axis. Plot false positives as red circles. Plot false negatives as green 'x' markers.

figure;
swarmchart(tp.Label,tp.Area,[],"k",XJitter="none",...
    LineWidth=1,SizeData=100,MarkerEdgeAlpha=0.4,DisplayName="True Positive (TP)")
hold on
swarmchart(fp.Label,fp.Area,[],"r",XJitter="none",...
    LineWidth=1.5,SizeData=100,DisplayName="False Positive (FP)")
swarmchart(fn.Label,fn.Area,[],"gx",XJitter="none",...
    LineWidth=1.5,SizeData=100,DisplayName="False Negative (FN)")
hold off
ylabel("Object Size")
xlabel("Class Name")
grid on
legend()

Figure contains an axes object. The axes object with xlabel Class Name, ylabel Object Size contains 3 objects of type scatter. These objects represent True Positive (TP), False Positive (FP), False Negative (FN).

[1] Adhikari, Bishwo; Peltomaki, Jukka; Huttunen, Heikki. (2019). Indoor Object Detection Dataset [Data set]. 7th European Workshop on Visual Information Processing 2018 (EUVIP), Tampere, Finland.

Open Live Script

Load precomputed detection results and ground truth labels. The ground truth data is a subset of the Indoor Object Detection Dataset created by Bishwo Adhikari [1]. The MAT files containing the precomputed results and annotations are attached to this example as supporting files.

load("yolov2IndoorObjectDetectorAnnotation100.mat");
load("yolov2IndoorObjectDetectorResults100.mat");

Evaluate Object Detection Results

Evaluate the object detection results against the ground truth using the evaluateObjectDetection function. Specify three overlap thresholds for the evaluation: 0.5, 0.75, and 0.9.

evalMetrics = evaluateObjectDetection(results,blds,[0.5 0.75 0.9]);

Compute Per-Image Metrics for Subset of Images

Compute the per-image metrics for a subset of images, and display the results.

imageResultsSubset = imageMetrics(evalMetrics,1:2:10); 
disp(imageResultsSubset)
    ImageIndex    NumPredictedObjects    NumGroundTruthObjects    TP    FP    FN    Precision    Recall
    __________    ___________________    _____________________    __    __    __    _________    ______

        1                  2                       2              2     0     0            1       1   
        3                  1                       1              1     0     0            1       1   
        5                  3                       2              2     1     0      0.66667       1   
        7                  2                       2              2     0     0            1       1   
        9                  2                       2              2     0     0            1       1

Compute Image Metrics at Specific Score Threshold

Specify a higher score threshold using the ScoreThreshold name-value argument, and compute the per-image metrics. At a higher score threshold, low confidence score predictions are dropped, leading to which will lead to fewer detected objects and more misses, or false negatives (FNs):

imageResultsThresh = imageMetrics(evalMetrics,1:2:10,ScoreThreshold=0.7);
disp(imageResultsThresh)
    ImageIndex    NumPredictedObjects    NumGroundTruthObjects    TP    FP    FN    Precision    Recall
    __________    ___________________    _____________________    __    __    __    _________    ______

        1                  2                       2              2     0     0         1           1  
        3                  1                       1              1     0     0         1           1  
        5                  1                       2              1     0     1         1         0.5  
        7                  1                       2              1     0     1         1         0.5  
        9                  1                       2              1     0     1         1         0.5

Compute Image Metrics at Specific Overlap Threshold

Specify a higher overlap threshold using the OverlapThreshold name-value argument, and compute the per-image metrics. The OverlapThreshold value must be one of the overlap thresholds at which the detection results were evaluated using the evaluateObjectDetection function.

imageResultsThresh = imageMetrics(evalMetrics,1:2:10,ScoreThreshold=0.7,OverlapThreshold=0.75);
disp(imageResultsThresh)
    ImageIndex    NumPredictedObjects    NumGroundTruthObjects    TP    FP    FN    Precision    Recall
    __________    ___________________    _____________________    __    __    __    _________    ______

        1                  2                       2              2     0     0         1           1  
        3                  1                       1              1     0     0         1           1  
        5                  1                       2              0     1     2         0           0  
        7                  1                       2              1     0     1         1         0.5  
        9                  1                       2              1     0     1         1         0.5

When you specify a higher overlap threshold, the detector misses more detections (more FNs).

[1] Adhikari, Bishwo; Peltomaki, Jukka; Huttunen, Heikki. (2019). Indoor Object Detection Dataset [Data set]. 7th European Workshop on Visual Information Processing 2018 (EUVIP), Tampere, Finland.

Open Live Script

Load precomputed detection results and ground truth labels. The ground truth data is a subset of the Indoor Object Detection Dataset created by Bishwo Adhikari [1]. The MAT files containing the precomputed results and annotations are attached to this example as supporting files.

load("yolov2IndoorObjectDetectorAnnotation100.mat");
load("yolov2IndoorObjectDetectorResults100.mat");

Evaluate Object Detection Results

Evaluate the object detection results against the ground truth using the evaluateObjectDetection function. Specify three overlap thresholds for the evaluation: 0.5, 0.75, and 0.9.

evalMetrics = evaluateObjectDetection(results,blds,[0.5 0.75 0.9]);

Identify Challenging Images Across All Classes

To identify images where the object detector performs poorly across all classes, analyze per-image detector performance using the false positive (FP) and false negative (FN) counts. First, compute the per-image metrics.

[summaryError,predsError,truthsError] = imageMetrics(evalMetrics,SortMetric="FP+FN",SortOrder="descend");

Select the 20 images with the highest number of incorrect detections, or the total number of FNs and FPs.

summaryError = summaryError(1:20,:);
predsError = predsError(1:20);
truthsError = truthsError(1:20);

Display the computed per-image metrics for the five images with the most incorrect detections.

disp(summaryError(1:5,:))
    ImageIndex    NumPredictedObjects    NumGroundTruthObjects    TP    FP    FN    Precision    Recall 
    __________    ___________________    _____________________    __    __    __    _________    _______

         4                 0                       2              0     0     2          NaN           0
        39                 4                       4              3     1     1         0.75        0.75
        66                 0                       2              0     0     2          NaN           0
        82                 1                       3              1     0     2            1     0.33333
         5                 3                       2              2     1     0      0.66667           1

Display the object predictions for the image with the index 39. The third "chair" class detection for this image is a FP, with its IsMatched column field set to false.

disp(predsError{2})
    ImageIndex    Label     Score     OverlapScore    IsMatched    MatchedTruthLabel    MatchedTruthIndex    Area             BoundingBox         
    __________    _____    _______    ____________    _________    _________________    _________________    _____    ____________________________

        39        chair    0.87222      0.74144         true          chair                     2            85352     468      61     227     376
        39        chair    0.86277      0.70789         true          chair                     1            98304     656      74     256     384
        39        chair    0.66098            0         false         <undefined>               0            28470    1018       1     146     195
        39        chair    0.62006      0.67055         true          chair                     4            55566     317      71     189     294

Identify Challenging Images for Specific Class

To identify images which are challenging for the object detector in detecting a specified class, analyze per-image detector performance using the FP and FN counts. To identify images where the detector struggled to identify objects of the "exit" class, specify the ClassNames name-value argument as "exit".

[summaryErrorClass,predsErrorClass,truthsErrorClass] = imageMetrics(evalMetrics,ClassNames="exit",OverlapThreshold=0.75,...
    SortMetric="FP+FN",SortOrder="descend");

Select the 20 images with the highest number of incorrect detections, or the total number of FNs and FPs.

summaryErrorClass = summaryErrorClass(1:20,:);
predsErrorClass = predsErrorClass(1:20);
truthsErrorClass = truthsErrorClass(1:20,:);

Display the computed per-image metrics for the five images with the most incorrect detections.

disp(summaryErrorClass(1:5,:))
    ImageIndex    NumPredictedObjects    NumGroundTruthObjects    TP    FP    FN    Precision    Recall
    __________    ___________________    _____________________    __    __    __    _________    ______

        20                 1                       1              0     1     1          0          0  
        26                 1                       1              0     1     1          0          0  
        47                 2                       2              1     1     1        0.5        0.5  
        55                 2                       2              1     1     1        0.5        0.5  
       100                 1                       1              0     1     1          0          0

Display the object predictions for the image with the image index 47, which has one FP and one FN. The second "exit" class detection for this image is a FP, with its IsMatched column field set to false.

disp(predsErrorClass{3})
    ImageIndex    Label     Score     OverlapScore    IsMatched    MatchedTruthLabel    MatchedTruthIndex    Area           BoundingBox       
    __________    _____    _______    ____________    _________    _________________    _________________    _____    ________________________

        47        exit     0.90114      0.91124         true          exit                      1            19902    464    221    186    107
        47        exit     0.68702      0.67807         false         <undefined>               0            29346    785    295    219    134

Display the ground truth information for the image. The second ground truth row for this image indicates a FN, with its IsMatched column field set to false.

disp(truthsErrorClass{3})
    ImageIndex    Label    IsMatched    Area           BoundingBox       
    __________    _____    _________    _____    ________________________

        47        exit       true       19890    463    221    195    102
        47        exit       false      27032    746    305    248    109

[1] Adhikari, Bishwo; Peltomaki, Jukka; Huttunen, Heikki. (2019). Indoor Object Detection Dataset [Data set]. 7th European Workshop on Visual Information Processing 2018 (EUVIP), Tampere, Finland.

Open Live Script

Load precomputed detection results and ground truth labels. The ground truth data is a subset of the Indoor Object Detection Dataset created by Bishwo Adhikari [1]. The MAT files containing the precomputed results and annotations are attached to this example as supporting files.

load("yolov2IndoorObjectDetectorAnnotation100.mat");
load("yolov2IndoorObjectDetectorResults100.mat");

Evaluate Object Detection Results

Evaluate the object detection results against the ground truth using the evaluateObjectDetection function. Specify three overlap thresholds for the evaluation: 0.5, 0.75, and 0.9.

evalMetrics = evaluateObjectDetection(results,blds,[0.5 0.75 0.9]);

Compute F1 Score Per Image

The F1 score is the harmonic mean of precision and recall, providing a balanced measure of classification performance. An F1 score of 1 indicates perfect precision and recall, while 0 indicates the worst performance.

Compute the image metrics over the data set.

[summaryError,predsError,truthsError] = imageMetrics(evalMetrics);

Display the first five rows of the per-image summary.

disp(summaryError(1:5,:))
    ImageIndex    NumPredictedObjects    NumGroundTruthObjects    TP    FP    FN    Precision    Recall
    __________    ___________________    _____________________    __    __    __    _________    ______

        1                  2                       2              2     0     0            1       1   
        2                  2                       2              2     0     0            1       1   
        3                  1                       1              1     0     0            1       1   
        4                  0                       2              0     0     2          NaN       0   
        5                  3                       2              2     1     0      0.66667       1

Note that precision is the ratio of the number of true positives and the total number of predicted positives​. The value of precision on the 4th row is NaN, since there are no predictions on this particular image, leading to a division by zero in the ratio calculation. Recall is the ratio of the number of true positives and the number of ground truth objects in that image. The value of recall for this image is zero, since no objects are detected, while there are two ground truth objects actually present in the image.

Compute the F1 score over the data set using the precision and recall values from the summaryError table.

precision = summaryError.Precision;
recall = summaryError.Recall;
f1Score = 2 * (precision .* recall) ./ (precision + recall);

Organize the precision, recall and F1 score values into a table.

prfTable = table(precision, recall, f1Score, VariableNames={'Precision', 'Recall', 'F1 Score'});

Display the precision, recall, and F1 score values for the first five images.

disp(prfTable(1:5,:))
    Precision    Recall    F1 Score
    _________    ______    ________

           1       1           1   
           1       1           1   
           1       1           1   
         NaN       0         NaN   
     0.66667       1         0.8

[1] Adhikari, Bishwo; Peltomaki, Jukka; Huttunen, Heikki. (2019). Indoor Object Detection Dataset [Data set]. 7th European Workshop on Visual Information Processing 2018 (EUVIP), Tampere, Finland.

Input Arguments

collapse all

Object detection metrics, specified as an objectDetectionMetrics object.

Indices of images for which to compute per-image object detection metrics, specified as a N-element numeric vector. Each element of the vector is a positive integer that represents an image index, and N is the number of images for which to compute metrics. By default, per-image metrics are returned for all images.

Name-Value Arguments

collapse all

Specify optional pairs of arguments as Name1=Value1,...,NameN=ValueN, where Name is the argument name and Value is the corresponding value. Name-value arguments must appear after other arguments, but the order of the pairs does not matter.

Example: imageMetrics(metrics,ClassNames=["person" "car"]) specifies to evaluate per-image metrics for the person and car classes.

Names of the classes for which to evaluate per-image metrics, specified as a string scalar or character vector. If multiple classes are specified, the function computes averaged metrics across the classes. By default, metrics will be computed for all classes stored in the ClassNames property of the objectDetectionMetrics object metrics.

Confidence score threshold, specified as a non-negative scalar. The imageMetrics function returns per-image metrics for detections with a confidence score below this threshold. The value of ScoreThreshold must be in the predicted confidence score range specified by the detectionResults argument of the evaluateObjectDetection function. Set the score threshold above the confidence threshold at which the detector was run.

Bounding box overlap threshold, specified as a numeric scalar or numeric vector. Each element of the numeric vector represents a box overlap threshold value over which the mean average precision is computed. By default, the overlap threshold is the first element of the OverlapThreshold property of the objectDetectionMetrics object metrics.

When the intersection over union (IoU) of the pixels in the ground truth bounding box and the predicted bounding box is equal to or greater than the overlap threshold, the detection is a match to the ground truth. The IoU is the number of pixels in the intersection of the bounding boxes divided by the number of pixels in the union of the bounding boxes.

Sorting metric that sorts the rows of the imageMetrics function output, specified as one of these options:

  • "none" — Do not sort.

  • "FP" — Sort by the number of false positive detections.

  • "FN" — Sort by the number of false negative detections.

  • "FP+FN" — Sort by the total number of false positive and false negative detections.

  • "Precision" — Sort by the precision value.

  • "Recall" — Sort by the recall value.

Tip

Use the SortMetric argument to identify images with the most False Positives or False Negatives. This can help determine the types of images for which the detector exhibits the lowest performance.

Sorting order for the rows of the imageMetrics function output, specified as one of these options:

  • "descend" — Sort in the descending order of the metric specified by the SortMetric name-value argument.

  • "ascend" — Sort in the descending order of the metric specified by the SortMetric name-value argument.

Output Arguments

collapse all

Metrics by area, returned as a table with N rows. N is the number of evaluated images. The columns are specified by these table variables.

  • ImageIndex — Index of the image, based on image order in the detectionResults argument of the evaluateObjectDetection function, returned as a non-negative integer.

  • NumPredictedObjects — The number of predicted objects in each image, returned as a positive integer.

  • NumGroundTruthObjects — Number of ground truth objects in each image, returned as a positive integer.

  • TP — Number of true positive detections, returned as a positive integer.

  • FP — Number of false positives detections, returned as a positive integer.

  • FN — Number of false negatives, returned as a positive integer.

  • Precision — Precision, returned as a positive numeric scalar. Precision is the ratio of the number of true positives (TP) and the total number of predicted positives​.

    Precision = TP / (TP + FP)

    FP is the number of false positives. Larger precision scores indicate that most detected objects match ground truth objects.

  • Recall — Recall, returned as a positive numeric scalar. Recall is the ratio of the number of true positives (TP) and the number of ground truth objects – the sum of true positives (TP) and false negatives (FN)​.

    Recall = TP / (TP + FN)

    FN is the number of false negatives. Larger recall scores indicate that more of ground truth objects are detected.

Per-image object predictions, returned as an N-by-1 cell array. N is the number of evaluated images.

Each element of the cell array is a table with P rows, where P is the number of predictions each image, and these columns.

  • ImageIndex — Index of the image, based on the image order in the detectionResults argument of the evaluateObjectDetection function, returned as a non-negative integer.

  • Label — Predicted class label, returned as a string scalar or character vector.

  • Score — Confidence score of the detector prediction, returned as a positive numeric scalar.

  • OverlapScore — Bounding box overlap score between the ground truth and detected bounding boxes, returned as a positive numeric scalar.

  • IsMatched — Logical that specifies if a predicted bounding box is a true positive or a false positive, returned as a numeric or logical 1 (true) or 0 (false), respectively.

  • MatchedTruthLabel — If the predicted bounding box is a true positive (matched to a ground truth), then MatchedTruthLabel specifies the class of the ground truth box, returned as a categorical array. Otherwise, it is undefined.

  • MatchedTruthIndex — If the predicted bounding box is a true positive (matched to a ground truth), then MatchedTruthIndex specifies the row index of the matched ground truth in the imageGroundTruth table, returned as a positive integer. Otherwise, it is 0.

  • Area — Area of the predicted bounding box, in pixels, returned as a positive integer.

  • BoundingBox — Bounding box, returned as a 4-element or 5-element numeric vector. For an axis-aligned bounding box, the vector is of the form [x y width height]. For a rotated bounding box, the vector is of the form [xcenter ycenter width height yaw]. x and y specify the upper-left corner of the rectangle. width and height specify the width and height of the rectangle, which are its lengths along the x- and y-axes, respectively.

If you specify "AOS" as an additional metric to compute using the AdditionalMetrics name-value argument of the evaluateObjectDetection function, each table contains these additional columns.

  • Orientation — The orientation angle of the predicted bounding box, returned as a non-negative numeric scalar. The angle specifies the clockwise rotation around the center of the bounding box. Units are in degrees.

  • OrientationDifference — The difference between the ground truth and predicted bounding box orientations, returned as a non-negative numeric scalar. Units are in degrees.

Per-image ground truth, returned as an N-by-1 cell array. N is the number of evaluated images.

Each element of the cell array is a table with M rows and these columns. M is the number of ground truth objects for a given image.

  • ImageIndex — Index of the image, based on the image order in the detectionResults argument of the evaluateObjectDetection function, returned as a non-negative integer.

  • Label — Predicted class labels of the ground truth, returned as a string scalar or character vector.

  • IsMatched — Logical that specifies if a prediction is matched to the ground truth or is a false negative, returned as a numeric or logical 1 (true) or 0 (false), respectively.

  • Area — Area of the ground truth bounding box, in pixels, returned as a positive integer.

  • BoundingBox — Bounding box, returned as a 4-element or 5-element numeric vector. For an axis-aligned bounding box, the vector is of the form [x y width height]. For a rotated bounding box, the vector is of the form [xcenter ycenter width height yaw]. x and y specify the upper-left corner of the rectangle. width and height specify the width and height of the rectangle, which are its lengths along the x- and y-axes, respectively.

If you specify "AOS" as an additional metric to compute using the AdditionalMetrics name-value argument of the evaluateObjectDetection function, each table contains this additional column.

  • Orientation — The orientation angle of the ground truth bounding box, returned as a non-negative numeric scalar. The angle specifies the clockwise rotation around the center of the bounding box. Units are in degrees.

Version History

Introduced in R2026a

See Also

| |

Topics