Hi everyone,
I’m running RT-DETR with a ZED camera on Jetson AGX Orin using the default Isaac ROS weights/config.
The ZED is set to 30 FPS, but when I run the detection pipeline, the image stream drops to around 11–14 FPS.
Is this expected for RT-DETR on AGX Orin?
Is the bottleneck likely the model inference, ZED pipeline, ROS queues/backpressure, image conversion, or GPU/CPU memory bandwidth?
What confuses me is: if inference is slow, why does the ZED camera stream FPS also drop? Does this mean the whole ROS pipeline is being bottlenecked?
Any profiling tips would be appreciated — TensorRT inference time, ZED node FPS, CPU/GPU usage, etc.