Secondray gie for full frame

• Hardware Platform Nvidia Tesla T4
• DeepStream Version 5.1

• TensorRT Version
• NVIDIA GPU Driver Version (valid for GPU only) 460.32…03
• Issue Type( questions, new requirements, bugs) question
• How to reproduce the issue ? (This is for bugs. Including which sample app is using, the configuration files content, the command line used and other details for reproducing)
use primary-gie and secondary-gie with different detector models.

Hi can i run secondary-gie as detector on full frame? (keeping primary also on full frame)
and can i get tracker ID for secondary-gie detections? (My ans - if class ID is different from primary detections, then i can get tracker ID for secondary detections also…Is this correct?)

I have seen back-to-back detector, i want to implement in deepstream-app.c. So may be simple parameters can do that in config file?

Then there will be 2 independent detectors, right?
Can you share your pipeline with us?

Yes, there are 2 detectors (person and face).

Here is my pipeline-

  1. Object detection as primary.
  2. Face detection as secondary.

There are 3 methods i tried-

  1. 2 instances of deepstream with 2 config files with separate detectors. But there is cost of decoder as for some sources, same frame has to be decoded in both the instances.

  2. back-to-back detector approach, but there secondary detections are depending on primary detected objects. I just wanted to know if i can run secondary detection on full frame?

  3. 3rd approach is i can use gst-dsexample, and get the frame with get_converted_mat API and do the face detection over there. But get_converted_mat API utilise CPU and my overall performance reduced. Is there method to get frame in gst-dsexample.cpp with out CPU involvement? like get frame from GPU directly?


Yeah, we can

Please tell how can i do that?

currently i am a bit confused. Like does secondary detection runs on every detected primary object? or primary-gie-id frame_meta(operate-on-gie-id)? as if it runs on primary-gie-id frame_meta that means it runs in full frame.

Still FPS goes bizarre when i use both as back-to-back detector. You can refer my post-

You can change the process-mode of sgie to make it work on full frame

I tried but its not supported i guess-
** WARN: <parse_gie:1373>: Unknown key ‘process-mode’ for group [secondary-gie]

hi, i updated process-mode=1 in property group of secondary-gie.
But still no improvement on overall performance.

4 streams with yolo_tiny gives 250 FPS each
4 streams with centerface only gives 200 FPS each.
4 streams with both model (as back-to-back detector yolo_tiny as primary-gie and centerface as secondary-gie) gives 30 FPS each. This is drastically low FPS.

Hey, I think this topic is talking about sgie for full frame. Please create a new topic if you still have perf issue.