image size of positive image set for training

We are using the trainCascadeObjectDetector( Current_Detector, imageData, negativeFolder, ... 'FalseAlarmRate', 0.5, ... 'TruePositiveRate', 0.995, ... 'NumCascadeStages', 10, ... 'FeatureType', 'Haar');

to train a cascade classifier.

Attached is a sample of a positive image for training (tropicana_cropped.png). I have also attached the test image (test_image.jpg).

Question: Should the sizes of the positive images (for training) be the same as in the objects-of-interest in the test image? For example, should the image of the "tropicana_cropped.png" be reduced in size to be very similar to the "Tropicana" in the test image before Training?

Thanks, Ed

Answer 1

[Tohru Kikawada]

You don't need to care the size of images. This is because vision.CascadeObjectDetector performs multiscale object detection on the input image. See this link for details.