Inference Using run_tfilte.py : Getting low score prediction also some objects are still not detected ?

[Kapil-23]

Generate TF-Lite model using below script :

!python model_inspect.py --runmode=saved_model --model_name=efficientdet-d1 \ --ckpt_path=AUTO_ML-50-30k --saved_model_dir=savedmodeldir-50-30k \ --tflite_path=efficientdet-d1-TF-50EP-30k.tflite \ --hparams=voc_config.yaml

I have trained efficientdet d1 (640x640) by setting --num_epochs=50 epochs. In some images model is able to detect objects. But In most of the other images model is not performing well. Also accuracy on detection remains near by to 70 %. Is there any parameter need to be included while converting the model or any preprocessing need to be done ?

One more query: At the time of inference i am manually resizing the image by keeping aspect ratio and pasting the image to 640x640 canvas by gray padding. Further i feed this image to model. FYI Model : Input : [1,640,640,3] Ouput : [1,100,7] Is this the correct way ?