isaac_ros_yolov8

Source code on GitHub.

Quickstart

1. Set up your development environment by following the instructions here.

2. Clone isaac_ros_common and this repository under ${ISAAC_ROS_WS}/src.

   cd ${ISAAC_ROS_WS}/src
   

   git clone https://github.com/NVIDIA-ISAAC-ROS/isaac_ros_common.git
   

git clone https://github.com/NVIDIA-ISAAC-ROS/isaac_ros_object_detection.git

3. Launch the Docker container using the run_dev.sh script:

   cd ${ISAAC_ROS_WS}/src/isaac_ros_common && \
     ./scripts/run_dev.sh
   

4. Install this package’s dependencies.

   sudo apt-get install -y ros-humble-isaac-ros-yolov8 ros-humble-isaac-ros-tensor-rt ros-humble-isaac-ros-dnn-image-encoder
   

5. Download the model of your choice from Ultralytics YOLOv8.

   Warning

   This step must be performed on x86_64.

   For this example, we use YOLOv8s.

   cd /tmp && \
      wget https://github.com/ultralytics/assets/releases/download/v0.0.0/yolov8s.pt
   

6. Export to ONNX following instructions here.

   Warning

   This step must be performed on x86_64.

   Arguments can be specified for FP16 quantization during this step. This ONNX model is converted to a TensorRT engine file and used with the Isaac ROS TensorRT node for inference. You can use netron to visualize the ONNX model and note input and output names and dimensions.

   This can be done by first installing ultralytics and onnx via pip:

   pip3 install ultralytics
   pip3 install onnx
   

   Afterwards, convert the model from a .pt file to a .onnx model using ultralytics. This can be done by running:

   cd /tmp && \
      python3
   

   Then within python3, export the model:

   >> from ultralytics import YOLO
   >> model = YOLO('yolov8s.pt')
   >> model.export(format='onnx')
   

   If you are planning on using Jetson, copy the generated .onnx model into the Jetson, and then copy it over into aarch64 Docker container.

   We will assume that you already performed the transfer of the model onto the Jetson in the directory ~/Downloads.

   Enter the Docker container in Jetson:

   cd ${ISAAC_ROS_WS}/src/isaac_ros_common && \
      ./scripts/run_dev.sh
   

   Make a directory called /tmp in Jetson:

   mkdir -p /tmp
   

   Outside the container, copy the generated onnx model:

   cd ~/Downloads && \
   docker cp yolov8s.onnx isaac_ros_dev-aarch64-container:/tmp
   

7. Run the following launch file to perform YOLOv8 detection using TensorRT:

cd /workspaces/isaac_ros-dev && \
   ros2 launch isaac_ros_yolov8 isaac_ros_yolov8_visualize.launch.py model_file_path:=/tmp/yolov8s.onnx engine_file_path:=/tmp/yolov8s.plan input_binding_names:=['images'] output_binding_names:=['output0'] network_image_width:=640 network_image_height:=640 force_engine_update:=False image_mean:=[0.0,0.0,0.0] image_stddev:=[1.0,1.0,1.0] input_image_width:=640 input_image_height:=640 confidence_threshold:=0.25 nms_threshold:=0.45

Note

A method of running the detections without visualization is available by using the yolov8_tensor_rt.launch.py launch file in this package.

The output bounding boxes here will be published on the /detections/output topic.

8. Run the image publisher node to publish input images for inference. A sample image (people_cycles.jpg) is provided here:

   cd /workspaces/isaac_ros-dev/src/isaac_ros_object_detection && \
      ros2 run image_publisher image_publisher_node resources/people_cycles.jpg --ros-args --remap /image_raw:=/image
   

9. Visualize and validate the output of the package in the rqt_image_view window. Inside the RQT image window that should have popped up, the output should look like:

API

Usage

ros2 launch isaac_ros_yolov8 yolov8_tensor_rt.launch.py model_file_path:=<model_file_path> engine_file_path:=<engine_file_path> input_binding_names:=<input_binding_names> output_binding_names:=<output_binding_names> network_image_width:=<network_image_width> network_image_height:=<network_image_height> force_engine_update:=<force_engine_update> image_mean:=<image_mean> image_stddev:=<image_stddev>

ROS Parameters

ROS Parameter	Type	Default	Description
tensor_name	string	"output_tensor"	Name of the inferred output tensor published by the Managed NITROS Publisher. The decoder uses this name to get the output tensor.
confidence_threshold	float	0.25	Detection confidence threshold. Used to filter candidate detections during Non-Maximum Suppression (NMS).
nms_threshold	float	0.45	NMS IOU threshold.

ROS Topics Subscribed

ROS Topic	Interface	Description
tensor_sub	isaac_ros_tensor_list_interfaces/TensorList	Tensor list from the managed NITROS subscriber that represents the inferred aligned bounding boxes.

ROS Topics Published

ROS Topic	Interface	Description
detections_output	vision_msgs/Detection2DArray	Aligned image bounding boxes with detection class.