Model trained with rect=True about yolov8 HOT 4 CLOSED

Are you having problems with regular detection?
Can you please explain to me about rect=True and how I should adjust this repo?

from yolov8.

Hello. Thanks for the reply.

rect=True tells the system that when performing recognition on an image that is not square the source image should keep its original aspect ratio and should be "letterboxed" with black on the short dimension.

Image Processing with rect=True: When rect=True is set, images are resized to match the longer dimension set by imgsz, maintaining their aspect ratio. This means the shorter sides might not be exactly to imgsz, but will be scaled accordingly.

Padding and rect=True: If there's empty space after resizing due to aspect ratio differences, padding is applied to maintain uniformity in dimensionality across the dataset for batch processing. YOLO uses zero-padding (pads with zeros) to fill these areas without affecting the image content.

I feel like the required change would be in the ImageSelector but I am not sure.

Regards

RKM

from yolov8.

In the current state if KeepOriginalAspectRatio setted as false the image is stretched without keeping the original aspect ratio, to the input size of the model, if KeepOriginalAspectRatio setted as true then the image keeps in the original aspect ratio and black padding is added to the margins, is there anything else you are looking for?

from yolov8.

No. I am sorry, I did not understand its operation. In reading I thought it changed the results after inferencing, not the source image.

Thank you

from yolov8.