FaceLandmarks is slow on Jetson Nano

[darekdev]

Hi.

I am using this FaceLandmarks model Facial Landmarks Estimation | NVIDIA NGC . And there is information that it should give around ~115FPS on Jetson Nano with one face but I got only 5-10FPS.
I use this repository for testing: GitHub - NVIDIA-AI-IOT/deepstream_tao_apps: Sample apps to demonstrate how to deploy models trained with TAO on DeepStream

I am running some project with face detection and I got 60FPS with 60FPS real time input and that is awesome but want to add some additional feature. With all standards properties. For example when I remove from pipeline face landmarks element (SGIE) from deepstream_facelandmarks_app.cpp then I see that there is fast face detector working like a charm. What can cause that face landmarks model perform much slower then from ngc documentation?

MAXN mode is on.

Best regards,
Darek

[AastaLLL]

Hi,

The table is tested by the trtexec tool, which measures the pure inference time.
Would you mind checking the model with trtexec to see if you can get a similar performance first?

Please remember to fix the clock to the maximum with tegrastats and use the fp16 mode for inference.

Thanks.