DeepStream-Yolo-Pose

NVIDIA DeepStream SDK application for YOLO-Pose models

YOLO objetct detection models and other infos: https://github.com/marcoslucianops/DeepStream-Yolo

Important: I've changed the output logic to prevent the TensorRT to use the wrong output order. Please export the ONNX model with the new export file, generate the TensorRT engine again with the updated files, and use the new config_infer_primary file according to your model

Getting started

• Supported models
• Instructions
• YOLOv7-Pose usage
• YOLOv8-Pose usage
• YOLO-NAS-Pose usage
• NMS configuration
• Detection threshold configuration

Supported models

• YOLO-NAS-Pose
• YOLOv8-Pose
• YOLOv7-Pose

Instructions

1. Download the DeepStream-Yolo-Pose repo

git clone https://github.com/marcoslucianops/DeepStream-Yolo-Pose.git
cd DeepStream-Yolo-Pose

2. Compile the libs

Export the CUDA_VER env according to your DeepStream version and platform:

• DeepStream 6.3 on x86 platform

  export CUDA_VER=12.1

• DeepStream 6.2 on x86 platform

  export CUDA_VER=11.8

• DeepStream 6.1.1 on x86 platform

  export CUDA_VER=11.7

• DeepStream 6.1 on x86 platform

  export CUDA_VER=11.6

• DeepStream 6.0.1 / 6.0 on x86 platform

  export CUDA_VER=11.4

• DeepStream 6.3 / 6.2 / 6.1.1 / 6.1 on Jetson platform

  export CUDA_VER=11.4

• DeepStream 6.0.1 / 6.0 on Jetson platform

  export CUDA_VER=10.2

Compile the libs

make -C nvdsinfer_custom_impl_Yolo_pose
make

NOTE: To use the Python code, you need to install the DeepStream Python bindings.

Reference: https://github.com/NVIDIA-AI-IOT/deepstream_python_apps

• x86 platform:

  wget https://github.com/NVIDIA-AI-IOT/deepstream_python_apps/releases/download/v1.1.8/pyds-1.1.8-py3-none-linux_x86_64.whl
  pip3 install pyds-1.1.8-py3-none-linux_x86_64.whl

• Jetson platform:

  wget https://github.com/NVIDIA-AI-IOT/deepstream_python_apps/releases/download/v1.1.8/pyds-1.1.8-py3-none-linux_aarch64.whl
  pip3 install pyds-1.1.8-py3-none-linux_aarch64.whl

NOTE: It is recommended to use Python virtualenv.

NOTE: The steps above only work on DeepStream 6.3. For previous versions, please check the files on the NVIDIA-AI-IOT/deepstream_python_apps repo.

3. Run

• C code

  ./deepstream -s file:///opt/nvidia/deepstream/deepstream/samples/streams/sample_1080p_h264.mp4 -c config_infer_primary_yoloV8_pose.txt

• Python code

  python3 deepstream.py -s file:///opt/nvidia/deepstream/deepstream/samples/streams/sample_1080p_h264.mp4 -c config_infer_primary_yoloV8_pose.txt

NOTE: The TensorRT engine file may take a very long time to generate (sometimes more than 10 minutes).

NOTE: To change the source

-s file:// or rtsp:// or http://
--source file:// or rtsp:// or http://

NOTE: To change the config infer file (example for config_infer.txt file)

-c config_infer.txt
--config-infer config_infer.txt

NOTE: To change the nvstreammux batch-size (example for 2; default: 1)

-b 2
--streammux-batch-size 2

NOTE: To change the nvstreammux width (example for 1280; default: 1920)

-w 1280
--streammux-width 1280

NOTE: To change the nvstreammux height (example for 720; default: 1080)

-e 720
--streammux-height 720

NOTE: To change the GPU id (example for 1; default: 0)

-g 1
--gpu-id 1

NOTE: To change the FPS measurement interval (example for 10; default: 5)

-f 10
--fps-interval 10

NMS configuration

For now, the nms-iou-threshold is fixed to 0.45.

NOTE: Make sure to set cluster-mode=4 in the config_infer file.

Detection threshold configuration

[class-attrs-all]
pre-cluster-threshold=0.25
topk=300

My projects: https://www.youtube.com/MarcosLucianoTV