Problem reading the data for YOLO v3 object detection

I am trying to implement a YOLO v3 object detector by following the tutorial documentation (https://se.mathworks.com/help/vision/ug/object-detection-using-yolo-v3-deep-learning.html?searchHighlight=yolo%20v3&s_tid=srchtitle_yolo%20v3_2). I am trying to train the computer to detect the rear of car and I have used 126 images of car rears as training images. However I can't seem to get it right even after having followed the tutorial thoroughly. Below is the output I get when trying to use the model.

I fear that my problem is that I'm not using the ImageLabeler app correctly, and maybe I am loading the data incorrectly? Currenctly I am manually labeling the images with rectangles and I am exporting the Image Labeling session Label File as a variable in the format of a 'table' called gTruth and calling that my vehicleDataset. Is this incorrect? Below is the following code:

doTraining = false;

if ~doTraining
    preTrainedDetector = downloadPretrainedYOLOv3Detector();    
end

%data = load('vehicleDatasetGroundTruth.mat');
%vehicleDataset = data.vehicleDataset;

% Add the full path to the local vehicle data folder.
%vehicleDataset.imageFilename = fullfile(pwd, vehicleDataset.imageFilename);
vehicleDataset = gTruth;

Other than that I am doing the exact same as the documentation. If anyone know what my issue may be then I would be very happy to hear your solutions.

Answer 1

[T.Nikhil kumar]

Hey Daniel!

As far as I can understand from the information you have provided, the way you have followed to load the data is correct. Your gTruth table’s structure is the same as the vehicleDataset in the given example.

(https://www.mathworks.com/help/vision/ug/object-detection-using-yolo-v3-deep-learning.html#ObjectDetectionUsingYOLOV3DeepLearningExample-2).

Coming to the error in detections, I can understand that the problem with your final detection is undesired multiple bounding boxes .

I suggest you to try using the selectStrongestBboxMulticlass() function to lessen the number of detections.

(Link:https://www.mathworks.com/help/vision/ref/selectstrongestbboxmulticlass.html)

Here, try to decrease the overlap ratio threshold to get rid of bounding boxes above that threshold value. Refer to the following snippet to update the previous [bboxes,scores,labels] to [newBboxes, newScores, newLabels]

[newBboxes, newScores, newClassNames] = selectStrongestBboxMulticlass(bboxes, scores, labels ,'RatioType', 'Union', 'OverlapThreshold', 0.3);

For your reference: https://www.mathworks.com/matlabcentral/answers/1667069-object-detection-using-yolov3-results-yolo-insert-of-multiple-boxes-for-the-same-object#answer_952610?s_tid=prof_contriblnk