how can I apply DeepStream-Yolo to my own application code？

[shengyu27]

How can I customize the post-processing logic code? I haven't found any py files that can be used to write code. Can they only be executed using the command line?
Or how can I apply DeepStream-Yolo to my own application code?

[edwardnguyen1705]

Hi @shengyu27 ,

This repo is used to convert model to the format that DeepStream based application can use.

Correct model format: TRT, ONNX, TTL (NvidiaTao toolkit). If you put ONNX model in the pgie cfg file, then pgie plugin will convert to TRT (using either default func or custom func). You have to provide post-processing lib that contains parse bboxes nvdsparsebbox_Yolo.cpp.

$ deepstream-app -c deepstream_app_config.txt

deepstream-app is an C++ based application that was already built/compiled. You can write custom apps or use sample apps to test the converted models.

[shengyu27]

thanks，I have another question. I saw this website and it's actually an application written based on Python interface, right?

[edwardnguyen1705]

thanks，I have another question. I saw this website and it's actually an application written based on Python interface, right?

That is the Python binding of /opt/nvidia/deepstream/deepstream-7.0/sources/apps/sample_apps. If you would like to use Python, then you can use that repo. However, for stability and full features, you must opt for deepstream-app.

[shengyu27]

I am using /opt/nvidia/deepstream/deepstream-7.1/sources/deepstream_python_apps/apps/deepstream-rtsp-in-rtsp-out to modify the code based on the foundation to complete my business logic, but I have found many bugs.
Are you saying that using the deepstream-app and configuring a custom model will result in better performance and stability?

[edwardnguyen1705]

@shengyu27 ,

Correct. However, you must write your biz logic in a callback function (probe).

[shengyu27]

Can you provide some help and tell me where to edit my custom logic? Or in other words, where to modify the callback function？

[edwardnguyen1705]

In /opt/nvidia/deepstream/deepstream-7.0/sources/apps/sample_apps/deepstream-app/deepstream_app.c, you must see a func,

gboolean
create_pipeline (AppCtx * appCtx,
    bbox_generated_callback bbox_generated_post_analytics_cb,
    bbox_generated_callback all_bbox_generated_cb, perf_callback perf_cb,
    overlay_graphics_callback overlay_graphics_cb)

There are two cb funcs: bbox_generated_post_analytics_cb and all_bbox_generated_cb. You must write your code in either.

[shengyu27]

Thanks, I will try it in the future.

[shengyu27]

I have to sigh, it's much better than the Nvidia tutorial.

[marcoslucianops]

https://github.com/marcoslucianops/DeepStream-Yolo?tab=readme-ov-file#extract-metadata

The C/C++ sample codes are in /opt/nvidia/deepstream/deepstream/sources/apps/sample_apps