Jetson Nano trt_pose_hand performance

Hi,

I’ve been playing around with my Jetson Nano and I ran the NVIDIA-AI-IOT/trt_pose_hand: Real-time hand pose estimation and gesture classification using TensorRT (github.com) example with my CSI camera

I get around 8.2FPS

I compared this with a custom built mediapipe 0.8 that includes compiled OpenGL ES support and I modified the Hands - mediapipe (google.github.io) solution to use either the CPU or the GPU and I get around 13-14FPS with the CPU version and around 40FPS with the GPU version.

That certainly surprised me a bit to see the performance difference so big…

Does anyone else have some performance statistics that they can share or any ideas what the reason may be?

Cheers
Andreas

I’m running

  • Jetpack 4.6
  • OpenCV 4.5.3 compiled from source with CUDA support
  • mediapipe 0.8 compiled from source with OpenEL GPU support. I modified the hand_gesture solution to use the GPU (different graph, …)

Hi,

The model complexity in these two examples are quite different.

trt_pos_hand sample use OpenPose resnet18_baseline_att_224x224 model whose size is 85MB.
mediapipe use TF lite model which is just 1.98MB.

It’s expected that smaller model gives a better performance.
But it may have some accuracy drawback in a complicated scenario.

Thanks.