How to Bring In New Object-Detection ONNX Models for Detectnet.cpp

[xplanescientist]

The Hello AI World project is working properly, pytorch and torchvision are installed and operational, and the jetson-inference Image-Classification and Posenet work properly with ONNX-based networks. This is how it was done on my Orin Nano devkit with Jetpack 6.2.

However, none of the Caffe-based networks work because they are not supported in TensorRT 10+, so the Object-Detection tools are still unuseable.

I want to get the Object-Detection tools operating again, specifically detectnet.cpp. I have working ONNX models, but the matrix output is different (e.g., 84x8400) from those expected by detectnet.cpp, so bringing in models only with the --network flag won’t work.

Many C++ programs reside in the ~/jetson-inference/c folder. Which functions should be copied/modified to get new ONNX models imported?

[AastaLLL]

Hi,

You can find most of the TensoRT-related implementations in the tensorNet.cpp

c/tensorNet.cpp

TensorRT also has a document for API changes to help users migrate.

https://docs.nvidia.com/deeplearning/tensorrt/latest/api/migration-guide.html#c

Thanks.

[xplanescientist]

Upon reviewing the tensorNet.cpp file, it’s quite involved. Project time is very important.

I’m working on a robotics project with an elaborate feedback control system. Object-Detection with Low Latency is of utmost importance, so my project is implemented in C++. I was previously relying on the Object-Detection tool (detectnet) from the Hello AI World project. Everything worked on my old Jetson Nano and a new Orin Nano with Jetpack6.0, but everything broke in Jetpack6.2 because the Caffe networks are no longer supported in TensorRT10+.

Currently, was it the best C++ Object-Detection approach that works in Jetpack6.2 so that I may integrate it with my robotics project? I can go back to Jetpack6.0, but then I lose the latest updates including MAXN Super. Please advise.

[AastaLLL]

Hi,

It looks like you file a new topic for the C++ implementation.
Please check that topic for the suggestion.

Yolov11 Inferencing Example in C++ for Jetpack6.2

Has Nvidia developed a C++ Object Detection example with Yolov11 or v8? Most of the inferencing examples are done in Python, not C++. Ultralytics says they don’t have a C++ example. My robotics project requires very low latency, so I need to implement the Object Detection in C++. I’m thinking the quickest and cleanest approach is to use the existing detectnet project from Hello AI World. But I would first need to modify the Yolo inputs and outputs to match what is expected by detectnet (not…

Thanks.