Explain why have the situational corrupt frame of video stream decoder in pipeline deepstream app

Accelerated Computing Intelligent Video Analytics DeepStream SDK
, ,
1

Hello everyone,

I have a question, in normal with having a few streams (e.g less than 3) and running with only detector engine. The output frames of decoder are smooths.

But why when increased number of streams or add more detector engine to pipeline (I have 3 detectors engine continue and some classifier engine), then the output frames of decoders sometimes corrupted, specially happened on the objects are moving. And I said that with streams RTSP do not have problem in transmission.

Thank you very much!

• Hardware Platform (GPU)
• DeepStream Version 4
• TensorRT Version 6.0
• NVIDIA GPU Driver Version 440.100

3

See inline.

4

–> the output frames corrupted, frames after nvinfer, right?
Yes it is
–> problem mean no frame corruption?
right.

5

Can you specify which GPU you are using?

6

I used some GPUs model, include: T4, P4, P100, 1080i. All have that problems.

7

How about GPU utility when corruption happened?
nvidia-smi -l 1
nvidia-smi dmon -i “gpu id”

8

Here they are:

GPU-Util about 80%~90%

nvidia-smi dmon -i 1

# gpu   pwr gtemp mtemp    sm   mem   enc   dec  mclk  pclk
# Idx     W     C     C     %     %     %     %   MHz   MHz
    1   142    64     -    73    32     0    10  3615  1531
    1    74    65     -    93    51     0     5  3615  1531
    1   210    65     -    90    46     0     9  3615  1531
    1   137    65     -    75    36     0     7  3615  1531
    1   204    66     -    85    43     0     9  3615  1531
    1    66    64     -   100     0     0     0  3615  1531
    1   117    64     -    90    51     0     6  3615  1531
    1   114    65     -    96    48     0     0  3615  1531
    1   212    65     -    76    32     0    10  3615  1531
    1   108    65     -    85    41     0     7  3615  1531
    1   230    65     -    71    30     0     8  3615  1531
    1   202    65     -    94    43     0     6  3615  1531
    1   137    66     -    86    41     0     3  3615  1531
    1   131    65     -    71    32     0    13  3615  1531
    1    79    65     -    77    36     0    12  3615  1531
    1   124    65     -    81    41     0     7  3615  1531
    1   108    65     -    81    38     0     3  3615  1531
    1   204    65     -    86    38     0     3  3615  1531
    1   189    65     -    70    33     0     9  3615  1531
9

How you check the corruption of frames? using File type to encode output to files or rtsp streaming?

10

I encode output and save to jpeg file with (multifilesink gst-element) to view them.

11

Are you using deepstream reference app or gst command?
How could we repro your issue?

12

Hello, I’m using deepstream reference app.

You would to like trying to add as much detector engine continue (which typical heavy like yolo or ssd) as I descriptive and trying to run with some (increase 5~10 stream) rtsp sources.

13

How about this part?

14

It’s kind of simple implement of a functions in deepstream_sink_bin.c in pipeline of deepstream app. I shared with you in PM. Please take a look.

15

Hi,
Just see you run with deepstream 4.0, please move to deepstream 5.0, we have many changes and fixes in this version, please try it and report if you still have the issue, if issue still there, we will see there.

1 Like