Discrepancy between observed performance (1.4 Hz) and official benchmarks (4.6 QPS) for FoundationPose on Jetson Orin

I am running Isaac ROS FoundationPose on a Jetson Orin, but I am only achieving a pose estimation frequency of approximately 1.5 Hz. According to the official NVIDIA documentation, the performance is rated at 4.6 QPS.

I would like to clarify if the reported 4.6 QPS figure refers strictly to the forward inference latency of the refine_model and score_model. Does this benchmark exclude the computational overhead of the GXF pipeline, specifically the pre-processing (sampling and Nvdiffrast rendering) and post-processing (decoding) stages?

The following are the execution latencies for the most time-consuming stages in the pipeline:
Pose Sampling and Rendering: 420 ms
Refine Model Inference and Sync: 240 ms
Score Model Inference and Sync: 200 ms

Start with the official NVIDIA benchmark definition and the GXF pipeline stages named in the issue: Pose Sampling and Rendering, Refine Model Inference and Sync, and Score Model Inference and Sync. Compare the reported 420 ms, 240 ms, and 200 ms latencies to determine whether 4.6 QPS covers only model inference or the full pipeline; done means documenting the benchmark scope and explaining the observed 1.5 Hz result.