retrain SSDLite MobileDet detector for the Edge TPU using my own images

[B0ndo2]

Hi,
I am trying to follow this tutorial to retrain the same pretrained model with my own images (about 300) to improve detection with a security camera (the angle of the camera is affecting detection).
I captured the images and used LabelImg to annotate the images and using Pascal/VOC format. I don't know how to take this new dataset and use it to do transfer learning, any help would be appreciated

[djalusic]

TensorFlow now has a new runtime framework for on-device AI, LiteRT, so TFLite (and TFLite Model Maker) is no longer maintained.

If you try to use these tutorial notebooks in Google Colab, they would fail since the packages aren't supported from Python 3.10. If you want to run it somewhere else, make sure to set it up with Python 3.9. (see issue here).

It seems that the current alternative to TFLite Model Maker is MediaPipe Model Maker, check out this tutorial if you want to use TensorFlow. There is a specific method for loading dataset in PASCAL VOC format (Dataset.from_pascal_voc_folder), make sure to use it when running the code.

If you aren't tied to TensorFlow because of TPU hardware, I suggest also training a model in PyTorch like YOLO. I switched from TensorFlow to PyTorch several years ago, it's much more user friendly.

[B0ndo2]

Thank you for the great information. Unfortunately I have to use Coral TPU so TF is my only option.
I will look at the tutorial and will come back with questions as I am by no means an expert.