Why the detection result seen on Nano is much worse than that seen on PC?

Hello, Dear NVIDIA teams,

My hardware/software info is :
• Hardware Platform (Jetson Nano )
**• DeepStream Version 5.0
• JetPack Version (4.3)
• TensorRT Version (7.1.3)
• NVIDIA GPU Driver Version (valid for GPU only)

We implemented YOLOv5l network with TensorRT API, and trained out a weights file (jde.wts) with the official YOLOv5l, on both PC and Nano we ran our code to create YOLOv5l network and load the weights file jde.wts and serialize an engine file,and then do inference with some pictures after having deserialized the engine file, we were surprised to have seen the person detection result seen on Nano is much worse than that we saw on PC, could you please help us resolve this issue ? we don’t know why the detection result is worse on Nano that got on PC with the same code and the same weights file.

I paste the sample result below:
Detection result got on Nano:

Detection result got on PC:

The original image:


Which TensorRT version do you use for desktop?

By the way, it seems you are still using JetPack 4.3.
Would you mind to upgrade it to the latest and try it again?


Thanks for your feedback.
On PC, TensorRT7.2.1 is used, on Nano TensorRT7.1.3 is used, this is the only difference, but for Jetson platform, it looks like even the latest JetPack 4.5.1 still includes TensorRT7.1.3 , no newer version is seen, according to the description here: JetPack SDK | NVIDIA Developer

Is there any way to forcibly install a newer version of TensorRT on Jetson platform?


JetPack 4.5.1 is the latest software version currently.
Is it possible to share a simple reproducible source as well as the model with us?

We would like to reproduce this issue in our environment.
And check this with our internal team.


Looking at the example images you have provided: “PC image” has a different x/y ratio, looks a little bit stretched on the vertical axis. Is this intended? Do you use exactly same image preprocessing steps on both platforms?


Do you fix this issue already?
If not, could you share the source and model with us for preproducing?