Using pretrained model on images of arbitrary size

[Mona-ShZeinoddin]

Hello,

Congratulations on the amazing work!,

I have images of size 320x256, I would like to use Reloc3r to perform relative pose estimation between pairs of images. Given both the width and height of my images are divisible by the patch size, my understanding is that I should be able to use the pre-trained models, could you please confirm this? if so, which one of the Reloc3r-512 or Reloc3r-224 do you recommend? Would be very grateful for your insight.

[siyandong]

Hi, thank you for your interest!

You can try both, just check out the demo code wild_relpose.py. Your images will be cropped and resized to either 224 or 512 pixels. Remember to set the same resolution for both the model, checkpoint, and images.

In general, 512 is better than 224. Given that your resolution is <512, perhaps 224 would be better to avoid blurry from image upscaling.