Object detection demo - changing resolution does not affect fps

Hi all,
I’m experimenting with the Jetson TX2 (and Nano) using provided demos from jetson-inference github https://github.com/dusty-nv/jetson-inference. When playing around with imagenet-camera demo, I notice that the network FPS indicator seems to be not affected by changing the input resolution. For ex. it fluctuates around 50-60fps for TX2 and 40-45fps for Nano. The webcam I used can provide 1080p@30fps video. So here are my questions regarding this behavior:

  1. The network fps means inference speed of the network and is independent from input video fps, am i right?
  2. If it's true, then I guess the API must has done some resizing/scaling pre-process for all input video, right?
  3. Then can I turn that feature off? If yes, could you pls show me how? I would like to see how it's actual performance scaling through various input video resolution/rate.

Thanks so much!

Hi,
The gstreamer string is in gstCamera.cpp. Please go ahead take a look. You may print out mWidth and mHeight for confirmation.
We also suggest you try DeepStream SDK 4.0.1

That is correct, it shows the inference speed of the network (also you can see a further breakdown in the console output)

Correct, for example with SSD-Mobilenet-v2 COCO model, the network expects a 300x300 input. So all input video is downscaled to the resolution that the network expects.

You would need to re-train the network at different resolution to feed in the camera data directly, without resizing. And even then it requires some layout conversion (i.e. to planar NCHW with mean pixel subtraction applied).