Based on my understanding, the detector head in this repo (featuresToBoxes()) is run on CPU host. I wonder why it is implemented in such way because it is not a common practice due to longer inference time. Is it because edge TPU does not support all operations of detector head?
Based on my understanding, the detector head in this repo (featuresToBoxes()) is run on CPU host. I wonder why it is implemented in such way because it is not a common practice due to longer inference time. Is it because edge TPU does not support all operations of detector head?