Object Detection using Yolov3 results : yolo insert of multiple boxes for the same object

Hello Matlabers,

I am usong YOLO v3 on matlab 2021a to recognize objects using YOLOv3 detector pretrained on COCO dataset.

the detector initiation is performed as:

name = 'tiny-yolov3-coco';

detector = yolov3ObjectDetector(name);

after the detection , I insert the bounding boxes using insertObjectAnnotation function

my problem is that the detector shows multiple boxes for the same object as shown in the image bellow

I am using grayscale image with replicated channels as Image=[gray, gray, gray]

the detection part is performed as follow:

[bboxes,scores,labels] = detect(detector,Image);

detected = insertObjectAnnotation(Image,'rectangle',bboxes,strcat(string(labels),{' - '},num2str(scores)));

what could be the problem and what are the possible solutions?

regards,

Answer 1

[Vivek Akkala]

Hi,

You can use selectStrongestBboxMulticlass to suppress the detections. Try lowering the "OverlapThreshold" of selectStrongestBboxMulticlass function. Following is the reference code:

[updatedBboxes, updatedScores, updatedClassNames] = selectStrongestBboxMulticlass(bboxes, scores, labels ,...

'RatioType', 'Union', 'OverlapThreshold', 0.3);

bboxes, scores and lables in the above code can be obtained from the detect(detector,image).

[ahcen aliouat]

Hi vivek,

Thank you for the sugestition, it helps me.

I have tried it before and the problem was that it eliminate the two inserted boxes in many cases. Which significantly reduce the accuracy. but as you said , I have to lowerise the overlap threshold to get best performances.