Confusing about TensorBBox TensorPoint and img_size

I’m slightly confused as to where the original image size which is accessed via get_meta("img_size") in transforms like Resize is actually added to the TensorBBox object. I read a bunch of source code but couldn’t find it (searching the whole source for set_meta didnt turn up any results).
Clearly it works in some examples for examples in @muellerzr Retinanet and Keypoint regression examples, but if I just create a TensorBBox manually and try to resize it, it will fail with something like:
TensorBBox doesn't have an attribute "img_size".

Any pointers?

You can either create your TensorBBox with an img_size argument and pass the corresponding image size, or if you have your images as inputs, the PointScaler transform will automatically add that metada when it sees the image input.

1 Like

Ahh okay I understand now, so the PointScaler will be applied to the images first and just store them away and then add them to the Boxes and Points it sees later.

I feel this should be documented. It is documented :slight_smile:


It is in the vision.core notebook :wink: (with examples too)

Oh nooo :confused: it is documented, sorry for bothering you, I clearly need to study all the notebooks end to end…

It’s alright! There’s a lot of notebooks. But it’s cool that the library is built on them. Usually I’ll look at the notebook now before looking directly at the source code if I’m confused on anything :slight_smile: (just a helpful hint for navigating it, as it is a little different than how documentation can be)

It’s also nice you can then run those notebooks top down and play with the code they’re using too