Problem with recognition rate when using SSD network

We are trying to image recognition on the SSD network using Tensorflow.
In the PC environment, I recognized all the classes well.
I did the same test on Jetson TX2, so I could not recognize some classes at all.

If you know the differences between the PC and TX2 installation environment below, please let me know.

[ PC ]

  • OS : Linux (Ubuntu 16.04)
  • Tensorflow Ver : 1.11.0 (Tensorflow-GPU) (Graphic Card : GTX1070)
  • Python Ver : 3.6.8
  • OpenCV Ver : 3.4.2

[ Jetson TX2 ]

  • JetPack : 3.3
  • OS : Linux (Ubuntu 16.04)
  • Tensorflow Ver : 1.9.0 (Tensorflow-GPU)
  • Python Ver : 3.5.2
  • OpenCV Ver : 3.4.0

As you probably see, you are using a different tensorflow version on your PC and on TX2, so that can happen.
I recommend you to update jetpack (https://developer.nvidia.com/embedded/jetpack) and tensorflow (https://developer.nvidia.com/embedded/downloads#?search=tensorflow) on your TX2.
Which exactly classes you cannot recognize?

Unrecognized classes are images of the same kind as other classes, but they are slightly different in appearance.
There are a total of 10 classes, some of which are not recognized at all.
However, all of them can be recognized on the PC.

That’s surprising that you get different results on TX2 if the model is the same.

Could you send us the model, and the scripts you use so that we can repro.

As Marek pointed out, it would be good to also test with the latest Jetpack and TF.
https://docs.nvidia.com/deeplearning/frameworks/install-tf-jetsontx2/index.html#install