Yolov5: Constraints Regarding sizes of images

[GlitchBox]

❔Question

Can I have images of numerous different sizes in my training set?
What do I do if I have to run inference on images of variable sizes?

Additional context

[glenn-jocher]

@GlitchBox you can train and detect on any size or shape images you want, in any combination. I would recommend you do so at similar train and detect sizes for best results.

[GlitchBox]

@glenn-jocher When you say 'in any combination', what do you mean? I understand that I can train on images of any size. But does that mean I can have both 1024x1024 and 1280x1280 images in trainset? Or do I have to resize all the images to either of the sizes(1024 or 1280)?
In the case that I don't have to resize images, what will the argument 'img_size' be? 1024 or 1280?

[glenn-jocher]

@GlitchBox you can train and detect on any size or shape images you want

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[AmirHosseinCV]

I have a similar problem. 50% of images are about 1080p and 50% of them are about 360p and I can't resize them (Large objects in small images and small objects in large images). A solution is using batch_size=1 and do not resize them, but It's not good. Is there a simple solution or I should change the code?

[glenn-jocher]

@AmirIR train with default settings.

[AmirHosseinCV]

@AmirIR train with default settings.

It resizes all images to 640 (here)

UPDATE: I solved the problem by customizing the multi-scale training feature. Thanks a lot!