If this is not possible, a workaround could be achieved if albumentation returns me the "perceived" visibility after the transformation. In that way I could calculate the "real" visibility as the multiplication of the initial and the perceived visibilities.

I do not understand the question yet.

As I understand, you crop parts from the image and bounding boxes that are not 100% contained in the image get truncated, right? And this becomes an issue.

Or not?

Could you provide some code?

Hi @ternaus , thanks for the reply.

Yes, you are correct. They get truncated. Therefore their visibility is not 100% to start with.

Unfortunately I don't think providing code will make things any clearer, because this is more of an workflow problem. So let me give an hypothetical situation:

1 - Imagine my dataset have images 2000x2000 px.
2 - My model just works with images of 500x500.
3 - Since I need the full resolution to identify the objects, I should not shrink the images. What I do instead is to slice the full res image (2000x2000) into 16 non-overlapping 500x500 patches.
4 - Now, imagine that there is a 2000x2000 image with two objects. During the slicing process, one object gets cut in half. The other stays fully visible in a single patch.

Everything up to this point happens before training the model. Its dataset pre-processing and has nothing to do with Albumentations.

5 - Now I'll start training a model and use albumentations. Then comes the question: Given that I want to work with minimal visibility of 40%, which value of min_visibility should I give to albumentations?

• If I use 40%, the object that was cut in half may end up being only 20% visible (40% of 50%)
• If I use a higher value, that would be too conservative for the object that stayed fully visible.