How to best deal with input images with fixed aspect ratio different from 1 (and not resizing) about blitznet

[dvornikita]

The right way to do it is to change the parameters of the initial tiling in the file boxer.py. The class PriorBoxGrid generates the tiling for a square area. If you make the tiling of an area with aspect ratio 1/2 you will have the right number of boxes that match the outputs of the convolutions.

Also, you may not want to do that at all because during the data augmentation parts of an image are randomly cropped and rescaled to a square. Since there are no constraints on cropping aspect ratio, the final aspect ratio of the patch is different from the original image.
What I'm saying is that even if you change the aspect ratio of the full pipeline you will still train on warped patches. If you train without random crops, the performance if likely to drop for the reason of lacking data augmentation.

[revilokeb]

@dvornikita yes, to adjust the tiling does make more sense.

Thank you also for pointing out that data augmentation does random cropping and then rescaling to square. That seems ok for my case however since I need to attach further network heads to the trunk which will require that aspect ratio.

I will close this when I have adjusted the tiling

[dvornikita]

No problems.

[juliagviu]

Could you please @revilokeb explain deeper how you adjusted the tiling? I am working with 512x1024 images and need to solve the same issue, but I am not sure how to make initial tilings of ratio 1/2. Thanks very much!