General information regarding input image

[ttdd11]

Your work is amazing.

In terms of the input image, if we are using 608x608 pretrained weights, do you recommend we resize the image to 608x608, or pad and resize to 608 to 608 for better results?

From the .dll (that I've used previously) I don't recall that there was a pad before the resize. What do you think?

[AlexeyAB]

In terms of the input image, if we are using 608x608 pretrained weights, do you recommend we resize the image to 608x608, or pad and resize to 608 to 608 for better results?

No.

Any image will be automatically resized to the network size.

[ttdd11]

I'm using tensorRT where we have to manage the input more comprehensively.

The problem I see is that calling a resize on the image with larger aspect ratios leads to some stretching and distortion.

In that case do you recommend padding before resizing to avoid this?

[AlexeyAB]

It depends on which approach of resizing did you use for training: #232 (comment)

[ttdd11]

Just using the pretrained weights from your site!

[AlexeyAB]

So it also depends on your images.
If aspect ratio ~ 1:1 then just use resizing.
If more than 1:2 or 2:1 than use letter_box (resizing with padding with keeping aspect ratio).