RSDK-10765 - Let mlvision resize to mlmodel's input size

[kharijarrett]

Hey, so this fix comes from a bug Nick ran into. Basically if someone uses a custom mlmodel module with a set input size different from the original image, we mess up the bounding box. Like, any absolute bbox coordinates that were read in were "normalized" using the original image and not the resized one. I think we've been so far getting lucky either with proportional bbox tensors or mlmodels that have [1 -1 -1 3] (variable) size input.

Here we just fix it by checking if we resized the image and using the appropriate image dims for normalization.

Tested via Python SDK with Nick's funky ncnn-ml module (0.1.1-rc1) and with our classic EffDet mlmodel setup. Nick's module used to not work now it does! And EffDet didn't break!

[HipsterBrown]

Since resizeW and resizeH could be set to the origW and origH, then just passing resizeW and resizeH in place of origW and origH in this method should fix the issue without any change to the method.

Suggested change

	boundingBoxes, err := ml.FormatDetectionOutputs(outNameMap, outMap, origW, origH, boxOrder, labels, wasResized, resizeW, resizeH)
	boundingBoxes, err := ml.FormatDetectionOutputs(outNameMap, outMap, resizeW, resizeH, boxOrder, labels)

[bhaney]

Create a unit test with a fake ML Model service that does this exact thing:

1. has a fixed input size which requires resize
2. outputs absolute bounding box coordinates relative to the resized image
3. Check that the bounding boxes from the vision service are what you expect

[bhaney]

Another thing -- public functions in the ml directory in RDK are used by ML on cloud inference, Do Not change these public functions, as it will introduce a breaking change.

Instead, feed in the resized dimensions into FormatDetectionOutputs's origW and origH

[bhaney]

You cannot change the signature of this public function - I believe it used by ML @tahiyasalam

[kharijarrett]

Sure. Made the fix without changing the method. Still working on writing the test.

[bhaney]

Thanks a lot! LGTM!