Human Reconstruction with RealSense

This tutorial demonstrates how to perform dynamic human reconstruction in nvblox using RealSense data. For more information on how human reconstruction works, see Technical Details.

Setup Isaac ROS Image Segmentation


If you are on a desktop machine, we recommend using the PeopleSemSegNet over the PeopleSemSegNet ShuffleSeg model that is provided in Isaac ROS Image Segmentation for better segmentation performance. If you are on a Jetson, using the PeopleSemSegNet ShuffleSeg model is recommended for overall inference speed and prediction accuracy trade-off.

The following steps show you how to run PeopleSemSegNet in ROS. See this document to run the PeopleSemSegNet ShuffleSeg network.

  1. Complete all steps of the Static Reconstruction with RealSense and make sure it is working.

  2. Clone the segmentation repository under ${ISAAC_ROS_WS}/src.

    cd ${ISAAC_ROS_WS}/src
    git clone
  3. Pull down a rosbag of sample data:

    cd ${ISAAC_ROS_WS}/src/isaac_ros_image_segmentation && \
        git lfs pull -X "" -I "resources/rosbags/"
  4. Launch the Docker container using the script:

    cd ${ISAAC_ROS_WS}/src/isaac_ros_common && \
  5. Inside the container, install isaac_ros_image_segmentation

sudo apt-get install -y ros-humble-isaac-ros-image-segmentation
  1. Download the PeopleSemSegNet ETLT file and theint8 inference mode cache file:

    mkdir -p /workspaces/isaac_ros-dev/models/peoplesemsegnet/1
    cd /workspaces/isaac_ros-dev/models/peoplesemsegnet
    wget ''
    wget ''
  2. Convert the ETLT file to a TensorRT plan file:

    /opt/nvidia/tao/tao-converter -k tlt_encode -d 3,544,960 -p input_1:0,1x3x544x960,1x3x544x960,1x3x544x960 -t int8 -c peoplesemsegnet_vanilla_unet_dynamic_etlt_int8.cache -e /workspaces/isaac_ros-dev/models/peoplesemsegnet/1/model.plan -o argmax_1 peoplesemsegnet_vanilla_unet_dynamic_etlt_int8_fp16.etlt
  3. Create the Triton configuration file called /workspaces/isaac_ros-dev/models/peoplesemsegnet/config.pbtxt with the following content:

    name: "peoplesemsegnet"
    platform: "tensorrt_plan"
    max_batch_size: 0
    input [
        name: "input_1:0"
        data_type: TYPE_FP32
        dims: [ 1, 3, 544, 960 ]
    output [
        name: "argmax_1"
        data_type: TYPE_INT32
        dims: [ 1, 544, 960, 1 ]
    version_policy: {
      specific {
        versions: [ 1 ]
  4. Inside the container, build and source the workspace:

    cd /workspaces/isaac_ros-dev && \
      colcon build --symlink-install && \
      source install/setup.bash
  5. (Optional) Run tests to verify complete and correct installation:

    colcon test --executor sequential
  6. Run the following launch file to get the ROS node running:

    ros2 launch isaac_ros_unet model_name:=peoplesemsegnet model_repository_paths:=['/workspaces/isaac_ros-dev/models'] input_binding_names:=['input_1:0'] output_binding_names:=['argmax_1'] network_output_type:='argmax'
  7. Open two other terminals, and enter the Docker container in both:

    cd ${ISAAC_ROS_WS}/src/isaac_ros_common && \
  8. Play the rosbag in one of the terminals:

    ros2 bag play -l src/isaac_ros_image_segmentation/resources/rosbags/unet_sample_data/

    And visualize the output in the other terminal:

    ros2 run rqt_image_view rqt_image_view
  9. Verify that the output looks similar to this image:

Example with RealSense Live Data

  1. Complete the Setup Isaac ROS Image Segmentation section above.

  2. Connect the RealSense device to your machine using a USB 3 cable/port.

  3. Run the ROS Docker container using the script:

    cd ${ISAAC_ROS_WS}/src/isaac_ros_common && \
  4. Source the workspace:

    source /workspaces/isaac_ros-dev/install/setup.bash
  5. Verify that the RealSense camera is connected by running realsense-viewer:

  6. If successful, run the launch file to spin up the example:

    ros2 launch nvblox_examples_bringup


If you want to restrict odometry to a 2D plane (for example, to run a robot in a flat environment), you can use the flatten_odometry_to_2d argument.

Example with RealSense Recorded Data

If you want to run the example on recorded data see RealSense Data Recording for Nvblox.


See RealSense Issues.