WTS Training tools #6
Description
We need to capture and further develop our system for training custom tflite object detection models for use on the Coral (via camml and in other tools) using labeled but unboxed images that are then boxed using MegaDetector to create the data for the object detector training. We call this process WTS because we are converting data from "Weakly To Strongly" labeled i.e. weakly labeled = only have a label for whole image, strongly labeled = have boxes and labels. While we have done some thinking about the validity of the assumptions inherent in this process, at some point we should enumerate, again examine, and capture those assumptions to help us develop this system in the best way.
However, this particular issue is about just collecting/building the bits and pieces that are necessary to generate data in this way and train from it.