Train with different image sizes

[MiguelAngeloMartinsRibeiro]

@dbolya I want to use your network to detect ships in the ocean. The problem is that I've three datasets and I want to train with images from all of them. The first one I created on blender and the image sizes are 550x550 pixels, so it's perfect to use yolact with the resnet101 backbone. The other one is 786x786 so it's still squared but I don't know if it makes a difference or not. The third one is more difficult because the resolution is 1920x1080. So what I'm asking is if I should make them all squared without caring about the size or should I resize them to 550x500 pixels.

Thank you for your attention

[st162053]

I am facing the same problem. Did it work with different datasets?