Disparity in FPS between /image_raw and /tag_detections using isaac_ros_apriltag_usb_cam.launch.py on Orin Nano

[TzuChiehHung]

I am running the default isaac_ros_apriltag_usb_cam.launch.py example on an NVIDIA Jetson Orin Nano to perform AprilTag detection using a USB camera.

I have observed a noticeable drop in the message rate (FPS) between the input image topic and the output detection topic.

Observed Rates in the Example Pipeline:

• Input Image Topic (/image_raw): Average rate is approximately 29 FPS.
• Output Detection Topic (/tag_detections): Average rate is approximately 18 FPS.

My Setup:

1. Hardware: NVIDIA Jetson Orin Nano
2. Launch Command: ros2 launch isaac_ros_apriltag isaac_ros_apriltag_usb_cam.launch.py
3. Screenshot: (Reference the attached screenshot showing ros2 topic hz results)

Benchmark Test Results

For comparison, I executed the official benchmark to establish the maximum possible throughput for the AprilTag node:

• Benchmark Command: launch_test $(ros2 pkg prefix isaac_ros_apriltag_benchmark)/share/isaac_ros_apriltag_benchmark/scripts/isaac_ros_apriltag_node.py
• Result: The standalone benchmark achieves a high FPS (consistent with the official figure of $\sim120$ FPS for the AprilTag Node/Graph at 720p on Orin Nano).

Question

1. Is the observed rate of $\sim18$ FPS on /tag_detections considered normal or expected when running the isaac_ros_apriltag_usb_cam.launch.py launch file?
2. Given that the standalone AprilTag node can achieve $\sim120$ FPS, what specific parameters or configuration aspects of the full pipeline (which includes the USB camera interface, the rectify node, and the AprilTag node) could be limiting the output rate to $\sim18$ FPS?
3. What specific steps can I take to debug or tune the pipeline configuration to achieve a /tag_detections FPS closer to the input image rate of 30 FPS?

[jaiveersinghNV]

Hi @TzuChiehHung ,

Thanks for confirming that the standalone Isaac ROS benchmark produces the expected result.

The "normal" rate of output detections is dependent on many factors. 18 FPS might be a reasonable number if there are other nodes running in parallel or if the system load is high for other reasons.

You can start by modifying the isaac_ros_apriltag_graph.py benchmark script to more closely resemble your target graph, and see if you can continue to observe ~120 fps numbers with those nodes. Note that our benchmarks run using pre-recorded rosbag data, not live sensor data from a camera.

If you're able to confirm that everything except for the USB camera node is capable of running at 120 FPS, then we know that the USB camera node is the main culprit. It's possible that loading the frames from the camera and publishing them is consuming enough CPU to bottleneck the system.

Also, note that ros2 topic hz is not always a good way to measure data frequency, since it can force the system to fully decode each message just to measure the frequency. Consider using NVIDIA's Greenwave Monitor instead for a more performant way to measure topic data.