0
$\begingroup$

In the paper in which DeepLabv3 is presented, the authors are mentioning that:

"For atrous convolution with large rates to be effective, large crop size is required; otherwise, the filter weights with large atrous rate are mostly applied to the padded zero region. We thus employ crop size to be 513 during both training and test on PASCAL VOC 2012 dataset"

What does this mean?
If for example I want to use the Cityscapes dataset for training, considering the idea mentioned by the authors, I should take 513x513 croppings for each example image that is originally 2048x1024?
Does the crop size refer to something else? Or if not, does it involve other steps/operations too?

I would be really grateful if somebody could clarify this concept and how it is applied during training and how it affects the inference.

Paper link: https://arxiv.org/pdf/1706.05587.pdf

Improve this question
$\endgroup$

1 Answer 1

Reset to default
1
$\begingroup$

In the meantime I found out that indeed the training of DeepLabv3 is performed in a patch-based manner which means the usage of a patch with a given crop size, that is extracted randomly from an example pair (input, target).

Regarding validation/testing, full-size example pairs (input, target) are used.

Improve this answer
$\endgroup$
1
  • $\begingroup$ As it’s currently written, your answer is unclear. Please edit to add additional details that will help others understand how this addresses the question asked. You can find more information on how to write good answers in the help center. $\endgroup$
    – Community Bot
    Commented May 10, 2023 at 13:19

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .