"argus_facedetectj" vs. "detectnet-camera facenet"

Autonomous Machines Jetson & Embedded Systems Jetson TX2
#1

Our result shows argus_facedetect can run >20 fps at 12M, but “detectnet-camera facenet” can only run ~5fps at 12M. Just wondering what will cause this difference?

The camera board is “LI-JETSON-KIT-IMX377” (from Leopard Imaging).

Thank you very much.

#2

Thanks a lot in advance.

#3

Hi xiong, We don’t have your camera board. Can it be reproduced with default camera board ov5693?

#4

Hi, DaneLLL. Thank you very much for your reply.

It can be reproduced with default board ov5693. But since the full resolution of ov5693 is 5M, the difference is not very obvious by human eyes (argus_facedetect runs ~30fps, detectnet-camera facenet runs ~25fps).

#5

Hi,

Suppose detectnet-camera will have lower performance.

Jetson_inference targets for demonstrating various deep learning use case (ex. classification, detection, segmentation).
Some optimization did’nt be applied for simpliciy.

There is two possible issue inside the pipeline:
1. A memory copy from GStreamer to CUDA
2. Multi-times overlay when rendering. More face will have lower performance.

However, argus is our deep learning pipeline specified for Argus and some optimization is done for it.
Thanks.

#6

Hi, AsstaLLL,

Thank you very much for your reply. It is very helpful for us.

#7