RTSP stream with image-meta-test

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Please provide complete information as applicable to your setup.

• Hardware Platform (Jetson / GPU) :- dGPU
• DeepStream Version :- 6.4
• Driver Version :- 535.183.01
• CUDA Version: :- 12.2

Input to pipeline :- RTSP stream from cameras connected to my local network.

I have a pipeline like this:-

pipeline
pipeline7861×407 192 KB

I have using probes from image-meta-test sample app, which encodes the frame and objects as JPEG.

When I run my pipeline with this probe, The stream that my display sink displays is not smooth, it keeps getting stuck and there is lot of flickering. I have attached a video for your reference.

when I remove this probe function, the stream works fine. I tested upto 16 streams at a time, all work fine (at around 25 FPS).

I have attached my pipeline code and a screen recording of what I’m getting.

forum.zip (5.7 MB)

3

Could you attach the loading of your GPU with command nvidia-smi dmon when encoding the image?

4

yes, please check this file.
forum_videos.zip (2.6 MB)

5

Not just nvidia-smi, you can run the nvidia-smi dmon and attach the log info. Thanks
Also could you tune the number of sources to see if it is related with the number of sources?

6

Tried it with 1,2,4,8 and 16 cameras with and without the obj_enc_api, the problem persists everytime when I keep the obj_enc_api enabled.

Also, I Have attached the nvidia-smi demon logs that you asked for. Thanks.

nvidia_smi_demon_logs.zip (2.4 KB)

7

Could you reproduce that with our sources\apps\sample_apps\deepstream-image-meta-test demo?
Also you can add your code based on this demo to reproduce this problem so that we can run that on our side.

8

sure. Will update you in some time.

9

app-for-nvidia.zip (66.1 MB)

models :- https://drive.google.com/drive/folders/1-0xuoIY9L9xJNiswt6kxSiAX39sFEAP9?usp=drive_link

I could not upload models as a part of the codebase because there is limit of 100MB on uploads. You will have to get the models from the above link and then put the models in app/ directory.

10

Thanks for your information. To narrow down the issue, I will try to reproduce that with our deepstream-image-meta-test first.

11

Sure. Thanks.

12

I have tried our deepstream-image-meta-test demo. When I commented out the probe function, the video display didn’t change significantly.
I also tried your project with 1 video source. It’s also have no significant changes when I set ObjEncAPI_enable to 0 and 1.
What model of GPU card are you using? Can you find a contrasting video source instead of the walk.mp4?

13

can you try it with RTSP?
I have 1650, 3060, 4090 and Quadro RTX 3000.
I observe the the problem with all the gpus with different number of sources. (eg, it takes more number of sources on 4090 for the problem to appear)

14

I cannot reproduce that either. The suspect point from your scenario is that the jpeg encoder and nvv4l2decoder may have conflict.
Since your project has more modules and a large amount of code, can you reproduce this problem with our deepstream-image-meta-test first?

15

Sure, I will update you.

16

I could not reproduce it Until I increased the number of sources to a very high number.

17

What sources did you use, and how many sources did you increase to reproduce the issue? And could you attach the result?

18

Yes. Sure. Would give you the results.

19

Source that I used :- Cameras on my local network giving me RTSP stream.
Number of Sources that I used :- 16

My Hardware :-

Dell Precision 7540,
OS :- Ubuntu 22.04.2 LTS
Memory :- 16GB
Processor :- Intel® Xeon(R) E-2276M CPU @ 2.80GHz × 12
Graphics Card :- NVIDIA Corporation TU106GLM [Quadro RTX 3000 Mobile / Max-Q] / Mesa Intel® UHD Graphics 630 (CFL GT2)

PFA a zip containing screen recording of my tests.

I used 16 cameras to reproduce the issue.

test_videos.zip (1.4 MB)

Do you want me to give you screen recording in a progressive manner ? I mean, test results on 8 cameras, then 12, 14 and so on ?

20

From the result video you attached, are you using our deepstream-image-meta-test without any change? And test2 is the result of commenting out the probe function, is that right?
Since test2 have no person, we can’t say if this is due to the commenting out reason.

21

I’m sorry for the delay.
Yes, I used the deepstream-image-meta-test sample app without any modification.
test_1 and test_2 both are results of deepstream-image-meta-test sample app as is, meaning no change was done.