mAP drops in the converted SSD mobilenet v2

[dsnsabari]

I followed all the steps, you have mentioned for converting TensorFlow SSD model v2 1.14 into TensorRT bin. I updated the buildengine.py file and graphsurgeon converter for new Jetpack support.

Jetpack version is 4.6

Do you the reason for the wrong bounding box?

[dsnsabari]

@jkjung-avt , I trained the SSD mobilnet v2 model on tf 14.0 and tested the model on tf 15.4 with Jetpack version 4.6

I trained my custom model with Object 1 and Object 2. The Object 2accuracy was too low for tensorRT. Could you tell me, how to improve the tensorRT accuracy?

	Object 1 count Accuracy	Object 2 Count Accuracy
TensorRT	41.48	94.6
TF frozen graph	86.807	93.2

[dsnsabari]

As per your suggestion, I had removed the FP16 from the build_engine.py file. But still, I am getting the same accuracy. There is no improvement
https://github.com/jkjung-avt/tensorrt_demos/issues/504#issuecomment-971066091