TLT with Deepstream

As below flow chart, I understood that primary detector(Gst-nvinfer) can divided two step as preprocessing and inference. Am I right? If so, Is metadata also made between preprocessing and inference?

Also, I want to use TLT to train new objects. As below TLT workflow, It seems like M5 is made and is put between training and inference steps. My question is that this metadata is put before preprocessing or is put between preprocessing and inference.


Additionally, in TLT workflow, does M2, M3, and M4 also means the metadata?