Here’s how you can phrase your question for a forum:
Question:
I have training images that are 1024 x 1024 pixels, and I’m training a YOLOv8 model, which requires input images to be 640 x 640 pixels. Roboflow offers several resize options, including “Stretch to,” “Fill (with center crop),” “Fit within,” and others. Which resize method would be the best option for resizing my images without distorting important features or losing crucial details for YOLOv8 training?
Hey @Rezwan-ul-alam_2011659042
Thanks for sending that in. For most cases, you would want to use the “Stretch to” resizing option to maximize the limited input space the model architecture (in this case, YOLOv8) can use to train on.
However, the main instances where you want to use the “Fill” or “Fit” processing steps are when the aspect ratio of the object (annotations) you are trying to detect matters for identifying it.
As an example, if you were trying to detect different cardboard box sizes (rectangle boxes, square boxes, etc) on images with different sizes and aspect ratios, a stretch processing step will mess up the lengths and widths (ie aspect ratio) in a way that will make it incredibly difficult for a model to learn the differences of each class. In this case, you would want to maintain the aspect ratio of the image (and annotations) by selecting a resize strategy that would not mess this up.
Should I use the “Fit within” method, then?
I am segmenting satellite images for buildings, trees, water, and farmland.
You would want to use stretch in your use case since I imagine that the dimensions of those objects don’t matter (ex: trees, no matter if they’re stretched one way or another, are still trees)
This topic was automatically closed 7 days after the last reply. New replies are no longer allowed.