Training procedure cropping and resizing for semantic segmentation

[ksagoog]

Hi,

Thanks for your great paper. For the semantic segmentation model on ADE20K, you state the following:

"""Images are resized to 520 pixels side length.
We use random horizontal flipping and random rescaling in
the range ∈ (0.5, 2.0) for data augmentation. We train on
square random crops of size 480."""

I feel I must not understand the procedure as randomly scaling a 520-pixel length image between the range (.5, 2.0) will result in some images of side-length less than your random crop size of 480. Could you please clarify the order of operations and any missing detail here? Thank you!

[ranftlr]

We used pytorch-encoding as our training framework. See here for the transform that is applied to the input during training: https://github.com/zhanghang1989/PyTorch-Encoding/blob/331ecdd5306104614cb414b16fbcd9d1a8d40e1e/encoding/datasets/base.py#L64

We used this class with base_size=520 and crop_size=480.

[ksagoog]

Thanks!