I recently trained an object detection model using Tensorflow 1.15 and the test result using Python is good. However, after converting it to .tflite format, running the model on Android showed a drastic drop in its performance.
Does the performance loss happen during model conversion to tf-lite? Is there anyway to avoid this loss of performance during conversion?
Reference:
Source of Training: https://github.com/tensorflow/models/tree/master/research/object_detection
Base Model for Transfer Learning: ssd_mobilenet_v1
Model Conversion: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/running_on_mobile_tensorflowlite.md
Python Test Script: https://github.com/tensorflow/models/blob/master/research/object_detection/object_detection_tutorial.ipynb
Android Demo App: https://github.com/tensorflow/examples/tree/master/lite/examples/object_detection
The first step I would do is to test it with a local Python Interpreter(after conversion); in that way, if you locally test with Python and the results are much poorer, then there is something wrong with the conversion. Normally, the post-training quantization should not drastically reduce the accuracy of your model, only by 2-3% in the worst case.
If the results are not bad when you feed images to your local Python Interpreter,(i.e. when you locally test your converted tf-lite model), then it means that there is a problem with the way you are feeding your input data on Android. Ensure that the exact same preprocessing steps are applied when feeding data to your images on your mobile app, like the ones during the training phase.