Image postprocessor

Open jonpsy opened this issue 4 years ago • 1 comments

Continuation of #679. @lu-wang-g

Dec 03 '21 09:12 jonpsy

Made substantial changes. To summarize:

a) I'm storing output_tensor_ in Postprocessor class. So that Postprocess() can access the corresponding output tensor and also find tensor_type. b) If output metadata doesn't have norm params, fallback to input metadata in Init(). This check is done via FindFirstNormProcessingUnit.

TODO: Currently, we pass first input tensor into BuildImageTensorSpecs, instead we should get the input_index similar to what was done here. That is out of the scope of this PR, we can create a follow up.

Also need to work on some docs, any suggestions?

Dec 19 '21 11:12 jonpsy