What image input size to use for training?

[haimat]

I have a custom dataset which I want to train docTR models on. These images all contain just a single line of text, hence they are about 2000 pixels wide and only 400 pixels in height. Now I saw that the default model input_size (W = H) is 1024 for the training script.

When I call a docTR model on an image, is this image then also resized to a fixed size?
What is the typical approach to find out what image size I should use during custom model training?
Last but not least, is it possible to train a model with different width and heigth sizes?

[felixT2K]

I have a custom dataset which I want to train docTR models on. These images all contain just a single line of text, hence they are about 2000 pixels wide and only 400 pixels in height. Now I saw that the default model input_size (W = H) is 1024 for the training script.

When I call a docTR model on an image, is this image then also resized to a fixed size? What is the typical approach to find out what image size I should use during custom model training? Last but not least, is it possible to train a model with different width and heigth sizes?

1. For inference it's also resized to 1024x1024 by keeping aspect ratio & symmetric padding (by default - the last two can be disabled only the size is fix)
2. See point 3 ^^ - in your case it would be 2048x2048 and I think this will slow down a lot and I expect the results to be also worse than with the default 1024 .. If you run the training with --show-samples how does it look still "natural" or "compressed" ?
3. It still is but was already planned to remove because it wouldn't work out of the box with the inference pipeline - and makes things a lot more complicated - and especially for detection the default 1024x1024 are really valid if we compare quality and inference latency

[haimat]

This is the output from --show-samples. I'd say they look "natural" to me.
So I guess I will stick to the default 1024 then.
Thanks for your quick response.