Sliding window inference overlap in 2D patches with 3 spatial dims (one of them being 1)

[YerePhy]

Question

• How does the monai.inferers.sliding_window_inference overlap parameter interact with patches 2D?

If the predictor eats patches of size (1, C, H, W, 1) and pops patches of the same size, there is no possible overlap across axis -1 or am I missing something?

[Nic-Ma]

Hi @YerePhy ,

Thanks for your interest here.
The overlap arg means: Amount of overlap between the sliding windows:
https://github.com/Project-MONAI/MONAI/blob/dev/monai/inferers/inferer.py#L100

Thanks.

[YerePhy]

Hi, thanks in advance for your answer

Yes, I've read that, but how you define overlap (please note that I'm talking about overlap here as a concept, not about overlap parameter) between sliding windows if their size is for example (H, W, 1)? I mean, the windows can overlap across axis 0 or 1 but not across axis 2, does monai.inferers.sliding_window_inference take that into account?

Look this example, supose the shape of the window is (H, W, 1):

[Nic-Ma]

Hi @YerePhy ,

For more details about the overlap & scan window interval, you can refer to:
https://github.com/Project-MONAI/MONAI/blob/dev/monai/inferers/utils.py#L166

Thanks.

[YerePhy]

Thanks @Nic-Ma,

The answer is yes, because interval is 1 if roi_size is 1.

It would be great to contribute to MONAI one of these days, but maybe I do not have enough experience yet.

[Nic-Ma]

Welcome your contribution!

Thanks.