Using DOPE at Different Image Resolutions

Overview

The DOPE network architecture, as outlined in the original paper, can receive input images of arbitrary size and subsequently produce output belief maps of the corresponding dimensions.

However, the ONNX format used to run this network on Triton or TensorRT is not as flexible, and an ONNX-exported model does NOT support arbitrary image sizes at inference time. Instead, the desired input image dimensions must be explicitly specified when preparing the ONNX file using the dope_converter.py script, as referenced in the quickstart.

Tutorial Walkthrough

  1. Complete until Run Launch File of the quickstart here.

  2. Under the Run Launch File section, run the dope_converter.py script with the two additional arguments row and col specifying the desired input image size:

    ros2 run isaac_ros_dope dope_converter.py --format onnx \
   --input ${ISAAC_ROS_WS}/isaac_ros_assets/models/dope/Ketchup.pth --output ${ISAAC_ROS_WS}/isaac_ros_assets/models/dope/Ketchup.onnx \
   --row 1080 --col 1920

  3. Continuing in the quickstart’s Run Launch File section, complete the data source-specific setup. Then, launch the data source-specific isaac_ros_examples.launch.py file with two additional arguments network_image_height and network_image_width specifying the desired input image size.

    For example:

    ros2 launch isaac_ros_examples isaac_ros_examples.launch.py <data source-specific arguments> ... \
   network_image_height:=1080 network_image_width:=1920  # Add these arguments to the launch command

  4. Continue with the rest of the quickstart. You should now be able to detect poses in images of your desired size.