A question about deepstream_parallel_inference_app

[lewis_bo]

**• Hardware Platform (Jetson / GPU) Jetson orin nano
**• DeepStream Version: 6.3
**• JetPack Version (valid for Jetson only) 5.1.2
**• TensorRT Version 8.5.2.2

I create my pipeline refer GitHub - NVIDIA-AI-IOT/deepstream_parallel_inference_app: A project demonstrating how to use nvmetamux to run multiple models in parallel., And the pipeline like this

I want to know if primary_gie_0_bin and primary_gie_1_bin synchroized?
For example, assume gie1_bin is run faster, gie1_bin process on 3th batch, but gie0_bin process on 1th batch

[Fiona.Chen]

All inferencing models work asynchronously. The metamux will synchronize the inferencing results.

[lewis_bo]

@Fiona.Chen

metamux

ths for your reply, and where can I get metamux usage documentation

[Fiona.Chen]

Yo can find the document here: deepstream_parallel_inference_app/tritonclient/sample/gst-plugins/gst-nvdsmetamux/README at master · NVIDIA-AI-IOT/deepstream_parallel_inference_app (github.com)

[lewis_bo]

Thanks.
I get a question with metamux. I set four source, only source0 can detect vehicle, but I get source1、source2、source3 data about vehicle from kafka.

And my metamux is

[lewis_bo]

But It works right when I replace source0 with source1 or source2、source3, I can’t get any data from kafka because no vehicle can be detected

[Fiona.Chen]

Please provide the app configuration together with metamux configuration

[lewis_bo]

config_metamux.txt (1.6 KB)
scs_pipeline.yml (6.0 KB)

[Fiona.Chen]

Your branch0 and branch1 share the same sources(0,1,2,3), why did you say “only source0 can detect vehicle,”??

Your model 1 and model 2 inference on all 4 sources according to your configuration. Please refer to NVIDIA-AI-IOT/deepstream_parallel_inference_app: A project demonstrating how to use nvmetamux to run multiple models in parallel. (github.com)

[lewis_bo]

I test two models, one is vehicle detection, the other is smoke or fire detection, source (0,1,2,3) is the test case I test vehicle detection, only source0 video has vehicles, source (1,2,3) video has no vehicles, so I say that “only source0 can detect vehicle”.

[Fiona.Chen]

So please post your sources_different_source.csv file too.

[lewis_bo]

sources_different_source.csv (269 Bytes)

[lewis_bo]

When I turned off the brand0 pre-processing and ran the vehicle detect model directly, it work right.
kafka has only source0 vehicle data.
My preprocess config:

[Fiona.Chen]

Why do you use nvpreprocess? Can you provide the graph png image, the graph image you post can not be viewed clearly.

[lewis_bo]

Because I only want to detecet the roi, not full frame.

[Fiona.Chen]

Can you upload the “PNG” file but not JPEG file? It is too small to view.

[lewis_bo]

How should I upload png, I upload the pipelien.png, size is about 2M, but I don’t know why it will become too small after uploaded

[lewis_bo]

[Fiona.Chen]

Can you enable the filesink and put the output video here for us to check how the bboxes looks like?

[lewis_bo]

Sorry so late to reply, I can’t save output video as mp4, but I recorded a video on my phone.