No dection on video with yolo models on deepstream sdk

Accelerated Computing Intelligent Video Analytics DeepStream SDK
, ,
1

• Hardware Platform (Jetson / GPU) Jetson Orin Nano
• DeepStream Version 6.3
• JetPack Version (valid for Jetson only) 5.1.5
• TensorRT Version 8.5.2
• NVIDIA GPU Driver Version (valid for GPU only) X
• Issue Type( questions, new requirements, bugs) questions

Hello,

I am currently attempting to convert a YOLO model to ONNX and apply it to the DeepStream SDK on the Jetson platform, following the official guidelines provided on the YOLO website.
(Ultralytics YOLO11 NVIDIA 에서 DeepStream SDK를 사용하는 Jetson 및 TensorRT)

While it works without issues on JetPack 6, I am encountering a problem with detection (bbox) recognition on JP5.1.5. Both YOLOv8 and YOLOv11 exhibit the same issue.

After conducting various tests, I found that the problem arises when using a CUDA-compatible version of PyTorch. Conversely, when using a non-CUDA-compatible version, the problem does not occur.

What I am looking for is a way to use a CUDA-compatible PyTorch on JP5.1.5 to convert the YOLO model to ONNX and apply it to the DeepStream SDK.

Could you please check if I might be missing anything ?

Thank you!

**Pytorch Version: ** 2.2.0
**Torchvision Version: ** 0.17.2
**Onnxruntime-gpu: ** 1.17.0

3

Could you try to use our repo yolo_deepstream?

4

Thank you for your reply.

However, the provided repository indicates that it supports DeepStream 7.0 and above, while I am using DS 6.3.

The repository you provided offers an ONNX file, but what I want is to implement the entire process of converting yolo.pt to yolo.onnx to TensorRT engine on Jetson. Importantly, the PyTorch I used is a CUDA-compatible version.

To provide additional context, I have installed the following versions of PyTorch that are CUDA-compatible:

  • 2.4.1 (custom build)
  • 2.2 (provided by the YOLO website)
  • 2.1 (provided by NVIDIA Jetson)
  • 2.0 (provided by NVIDIA)
  • 1.14.0 (provided by NVIDIA)

I have tested these versions but have faced failures. However, it worked on JP6, and I don’t understand why it doesn’t work on JP5, so I am inquiring about this.

5

OK. Then you can refer to DeepStream-Yolo, this project contains the transformation method and works with all the DeepStream versions.

6

Thank you for your response.

As a result, I solved it using yolo_deepstream

Thank you. Have a great day!