Object Detection Workshop#

This tutorial runs the OpenVINO™ object detection tutorial in a workshop, with inference on the CPU, Intel® GPU and Intel® NPU. The workshop installs the full Robotics AI Suite with the official robotics-installer, plus ros-${ROS_DISTRO}-object-detection-tutorial.

You will:

  1. Launch the workshop and connect the NPU and the desktop.

  2. Compare inference latency on the CPU, GPU and NPU.

  3. Watch detections in RViz on each device.

Prerequisites#

  • A host set up as described in Getting Started.

  • A local desktop session for the RViz steps.

Launch the workshop#

  1. Go to the workshop directory and launch it. The first launch takes 10 to 30 minutes:

    cd src/components/object-detection/workshop
    workshop launch object-detection
    
  2. Connect the NPU and the desktop. Neither connects automatically:

    workshop connect object-detection/robotics-ai-suite:npu :custom-device
    workshop connect object-detection/robotics-ai-suite:desktop
    
  3. Apply the NPU fix. It takes about 30 seconds and is needed for the NPU to publish detections; see Known issues:

    workshop run object-detection build-npu-fix
    

Measure inference latency#

Run OpenVINO™ benchmark_app with the tutorial’s model on each device:

workshop run object-detection benchmark

On an Intel® Core™ Ultra X7 358H processor, median latency is about 3.8 ms on the CPU, 0.8 ms on the GPU and 0.9 ms on the NPU. The action also reports NPU busy time.

To measure the rate of the full ROS 2 detection pipeline without opening a window, run:

workshop run object-detection benchmark-ros

It also prints which openvino_wrapper_lib is loaded, so you can confirm the NPU fix is in use.

View detections in RViz#

Run the tutorial on one device at a time:

workshop run object-detection ros-object-detection-gpu

RViz opens and shows the detections. Press Ctrl+C in the terminal to stop it, then try ros-object-detection-cpu or ros-object-detection-npu.

Actions#

Action

Description

build-npu-fix

Rebuilds openvino_wrapper_lib with the SSD NPU fix as an overlay in ~/ov_ws. Once built, the other actions use it automatically

ros-object-detection-cpu, -gpu, -npu

The tutorial launch with RViz, on the selected device

benchmark

benchmark_app latency on the CPU, GPU and NPU, plus NPU busy time

benchmark-ros

Headless detection pipeline rate on the CPU, GPU and NPU

Performance compared with the host#

The same benchmark_app run (OpenVINO™ 2025.3, latency hint, 10 s) on the host and in this workshop, alternated three times per device, gave 98% (CPU), 101% (GPU) and 101% (NPU) of the host’s throughput. The differences are within run-to-run variation: running in a workshop adds no measurable inference overhead.

Humble#

Use object-detection-humble in every command. benchmark and benchmark-ros work on the CPU and GPU. The NPU does not work on Humble, and build-npu-fix does not apply; see Known issues on Humble.

Known issues#

  • No NPU detections without the fix (Jazzy). With ros-jazzy-openvino-wrapper-lib 2025.3.0-0~0.11.0eci10, the NPU runs inference but the tutorial publishes no detections. The cause is the SSD preprocessing in ros2_openvino_toolkit, fixed upstream in commit f6e0f5d, which is not yet in a released package. build-npu-fix applies that commit. CPU and GPU are not affected.

  • A rebuild removes the fix. workshop refresh after an SDK change rebuilds the container and removes ~/ov_ws. Run build-npu-fix again afterwards.

Source#

The definitions, package list and patch are in src/components/object-detection/workshop/.