Skip to content

Latest commit

 

History

History
113 lines (94 loc) · 3.79 KB

README.md

File metadata and controls

113 lines (94 loc) · 3.79 KB
Raw

YOLOv8-for-Jetson-Orin

Before you go

This code is for Jetpack 5.1.3, so ENSURE that your jetson is using Jetpack 5.1.3

  • Jetpack 6.0 will work too but compatible with another version of Pytorch(2.1.0) & Torchvision(0.16.1)
  • So if you are in Jetpack 6.0, edit these versions in "startup.sh"

Version differences of main packages are described in appendix for someone in need

Virtual Environment User

If you are going to use this code in virtual environment, please check your OpenCV build informantion in your virtual environment(NOT HOST)

$ python3
$ >>> import cv2
$ >>> print(cv2.getBuildInformation())

From this information, you can run this code if only "GStreamer" is "YES"

  • If this is "NO", then you should use your host environment

Please check your host environment OpenCV too for sure

  • SDK will automatically installed OpenCV with GStreamer "YES"
  • But if it says "NO" in host environment, you may install "opencv-python" additionally so uninstall it
pip3 uninstall opencv-python

Start Up (DO NOT USE sudo FOR startup.sh)

Preparing environmet for running YOLOv8 in Jetson, using CSI Camera

OpenCV 4.5.4 (SDK already installed it for you)
GStreamer 1.16.3 (SDK already installed it for you)
Pytorch 1.13.0
Torchvision 0.14.1
ultralytics 8.1.37
onnx 1.16.0
numpy 1.21.3

Installation of Pytorch and Torchvision are from https://forums.developer.nvidia.com/t/pytorch-for-jetson/72048

DO NOT USE 'sudo' command

  • Since we build torchvision from source, if you execute startup.sh with sudo, torchvision will be accessible only with sudo command
sh startup.sh

Demo Setup Description

This code is using YOLOv8 model for demonstration

Model Setting

  • 80 classes to predict
  • Basically YOLOv8n is included in each directory (Model in FP16 and FP32 will be converted when you run it)
  • If you want another version like "YOLOv8m", "YOLOv8l" etc.. just replace it in every directory

Image Setting

Capture Width = 1920
Capture Height = 1080
Display Width = 640
Display Height = 360

Error Handling
For generality there are some "rm" and "export" command to cope with some errors may occur

  • Cannot allocate memory in static TLS block

For these error, we have to preload modules "libgstnvarguscamerasrc.so" and "libgstnvvidconv.so"

  • Stalled when building GStreamer pipeline

There are some cases that GStreamer register "libgstnvarguscamerasrc.so" to its BLACKLIST.
With this situation, demo.sh will just stunned with no error message which makes you suffer.
To solve this problem, we should remove GStreamer cache.

Demonstration

After running "startup.sh" please reboot your jetson for safety.

For demonstration, you should select model and framerate.

There are 3 models (PLEASE pass exact parser)

  • Original: Original YOLOv8 model from ultralytics package
  • TensorRT-FP16: TensorRT converted model with FP16
  • TensorRT-FP32: TensorRT converted model with FP32
  • Framerate: Recommanded about 15~25
sh demo.sh {MODEL_NAME} {FRAMERATE}

Appendix

Jetpack 5.1.3

Ubuntu: 20.04
Linux kernel: 5.10 LTS
OS: L4T 35.5.0
CUDA: 11.4.19
cuDNN: 8.6.0
TensorRT 8.5.2
VPI: 2.4
OpenCV: 4.5.4
GStreamer: 1.16.3
Python2: 2.7.18
Python3: 3.8.10

Jetpack 6.0

Ubuntu: 22.04
Linux kernel: 5.15 LTS
OS: L4T 36.2
CUDA: 12.2.12
cuDNN: 8.9.4
TensorRT 8.6.2
VPI: 3.0
OpenCV: 4.8.0
GStreamer: 1.16.3
Python2: 2.7.18
Python3: 3.10.12