Jetbot Object Detection Engine

Is it possible to use a custom engine for object detection?
I have a custom coco dataset but it’s .coco not .engine.
Any advice would be appreciated :)

Thank you!

1 Like

Hi,

Can you please tell me what Jetson product you are using? I can then move your topic to the correct category. This section is used only for reporting community issues and is not monitored by support staff.

Thanks,
Tom

1 Like

I responded in the other thread.
But it’s jetson nano 4gb.
Thank you Tom

Moving this topic to Jetson Nano.

Thanks,
Tom

1 Like

Sorry Tom, I can’t find the Jetson nano forum. Could you kindly send me a link?
Thank you

Hi,

Jetbot uses TensorRT as the inference engine.
So you will need to convert the model into a TensorRT engine first.

May I know which training frameworks do you use?
It seems that you are using coco dataset. But what is your model format?

Thanks.

I’m using roboflow to export them as a coco model. Went for coco as this was used as the example in object following.
I don’t believe I have a tensorRT option in roboflow. How do you recommend I convert it to a TensorRT engine?

Hi,

In general, you will need an ONNX format model to deploy with TensorRT.

It seems that roboflow doesn’t support converting COCO to ONNX directly.
You might try some intermediate format to see if it works instead.

For example, COCO - PyTorch - ONNX - TensorRT.

Thanks.

Thanks AastaLLL,
Would yolv5 pytorch onnx tensorrt work?
What Ive run into are version issues. For example dif. Tensorrt version for conversions to use in the jetbot sdk. Ive had to reflash because some installments upgraded some packages and were no longer compatible with the examples :(
Is there a guide I can follow for the way the object detection coco engine was created for the jetbot?

Hi,

The PyTorch - ONNX - TensorRT should work.
For the compatibility issue, do you have errors or logs can share with us?

Thanks.

Sorry not yet, still trying to figure out how to do it. I have built a dataset in roboflow and have trained a model with yolov5 so far. Will let you know if I run into issues converting to ONNX and TensorRT