Human Pose Estimation: NCS2 Slower Than CPU

[magoldsm]

Just purchased a NCS2.  Running the Open Model Zoo demo of Human Pose Estimation (human_pose_estimation_demo\python.  Using the 0005 model).  It runs faster on my Core i9-9980HK (29 fps) than on the NCS2 (8 fps).  It does not matter which model I use - all are 3-4x faster on CPU.

Is this normal?  Running OpenVINO 2021.4.689

[Iffa_Intel]

Greetings,

The performance on CPU is expected to be better compared to NCS2 since CPU has more powerful computing power.

NCS2 is just an accelerator device which would help in a certain situation, especially when additional computing power is required.

You may refer here for more results comparison:

https://docs.openvinotoolkit.org/latest/openvino_docs_performance_benchmarks_openvino.html

Besides, CPU required FP32 model format while NCS2 required FP16 model format. FP16 might have something called Quantization Error since it is squeezed from a full precision model which made it smaller in size. This would somewhat affect accuracy/performance.

You may refer here for further understanding: https://www.youtube.com/watch?v=RF8ypHyiKrY

Sincerely,

Iffa

[Iffa_Intel]

Greetings,

Intel will no longer monitor this thread since we have provided a solution. If you need any additional information from Intel, please submit a new question. 

Sincerely,

Iffa