Optimizing Grayscale Image Acquisition for VSLAM on Jetson AGX Orin

Software ROS2
1

Hello ZED Community,

I am currently integrating a ZED camera with NVIDIA’s Isaac ROS Visual SLAM (VSLAM) on a Jetson AGX Orin platform. As VSLAM requires grayscale image input, I am exploring the most efficient method to acquire these images.​

Options Considered:

  1. Subscribing to Native Grayscale Topics: The ZED ROS 2 wrapper publishes rectified grayscale images on topics such as /zed/zed_node/left/image_rect_gray and /zed/zed_node/right/image_rect_gray. Subscribing directly to these topics leverages the camera’s internal processing capabilities, potentially minimizing additional computational overhead.
  2. Utilizing NVIDIA’s Image Format Converter Nodes: Alternatively, NVIDIA’s isaac_ros_image_proc package offers GPU-accelerated nodes for image processing tasks, including format conversion. By subscribing to the color image topics (e.g., /zed/zed_node/left/image_rect_color), these nodes can convert the images to the desired grayscale format, utilizing the AGX Orin’s GPU resources.

Specific Questions:

  • Grayscale Image Generation: How are the grayscale images on the native topics generated? Are they produced directly by the ZED camera’s onboard processing, or does the ZED ROS 2 wrapper perform the conversion from color to grayscale?​

  • Performance Considerations: Which method is more efficient on the Jetson AGX Orin platform? Is it better to utilize the native grayscale topics, potentially reducing CPU load, or to employ NVIDIA’s GPU-accelerated image format converter nodes for grayscale conversion? Are there any benchmarks or experiences related to CPU/GPU utilization and latency for these approaches?​

I appreciate any insights or experiences you can share regarding the optimal approach for acquiring grayscale images for VSLAM on the Jetson AGX Orin.

Thank you!

2

Hello,

Thanks for your message.

Grayscale Image Generation : The wrapper gets the grayscale image directly from the SDK. It doesn’t perform the conversion from color to grayscale. The Grayscale calculation is performed on the GPU.
Performance Considerations: We do not have direct benchmarks comparing both methods.

As general best practices to use our ROS2 wrapper with the best performances, follow these guidelines : https://www.stereolabs.com/docs/ros2/150_dds_and_network_tuning.

In addition to the DDS tuning, ensure you are using MAXN mode on your Orin AGX. Pay attention to the use of the compressed topics with image_transport as described in the tutorial.

Best,
Rodolphe Perrin

Stereolabs Support

1 Like