Feature request: Support UInt16 inputs

[tayden]

It's not uncommon to have satellite imagery distributed in unsigned 16 bit format. It'd be great if this plugin would support arbitrary bit depth images instead of only uint8 types.

I image that this could probably happen automatically be reading the input image metadata, or that users could define this value in the inputs.

[bartoszptak]

Hi @tayden !

We can consider the feature in the next release. Could you share an example data, model, and preprocessing description? If the data is confidential, please use the email address listed in my profile.

[tayden]

I think the main thing that needs to happen to make this work is that this line needs to have a configurable max_pixel_value parameter that could be defined in the onnx metadata. This could be similar to how the standardization_mean and standardization_std values are defined.

qgis-plugin-deepness/src/deepness/processing/models/preprocessing_utils.py

Line 22 in 2deec30

return np.float32(tiles_batched * 1./255.)

I don't have any data I'm permitted to share. I was able to get this to work with a bit of a hack where I scaled up my normalization parameters up to compensate for [0,1] scaling that assumed a max pixel value of 255.

If I can find some time, I can work on adding a PR to accomplish this. We could just assume that the max pixel value is 255 unless a user specifies it as something else.

[tayden]

As an additional reference, what I'm suggesting is based on how Albumentations handles their standard normalization transform: https://explore.albumentations.ai/transform/Normalize