Postprocessor performed on Primary Inference Source Probe not allowing Secondary Inference to Process results

[abdo.babukr1]

• Hardware Platform (Jetson / GPU)
Jetson Orin AGX Dev Kit
RTXA4000, Driver Version 515.86.01 Dell XPS intel I7-10700 CPU
• DeepStream Version
6.1.1
• JetPack Version (valid for Jetson only)
5.0.2

We have a post processing library that works with a secondary inference, we are able to access the output tensor of the secondary inference in a source probe. We are using a python application.

The issue is the obj_meta.obj_user_meta_list shows as None when implementing the exact post processor on the primary inference source probe instead of using a postproc library.

The object meta gets placed correctly including the rect_params, confidence, classid, object_id, object label, text_params, bboxinfo, and unique_component_id.

On the frame level, the number of objects is set correctly, bInferDone is set to 1, and the source width/height are set correctly. We can access the frame and object meta data in the secondary source probe as shown:

obj_user_meta_list None, height 375.0 width 162.0 class_id 1
top 233.0 left 271.0
bbox info: left 271.0406188964844
bbox info: top 233.3119354248047
bbox info: width 162.89999389648438
bbox info: height 375.16741943359375

As you see the obj_user_meta_list is None, we know the models are working and the configuration files are correct because, as mentioned, we are able to extract the obj_user_meta_list from the post processor library implementation.

So my question is whats missing here? Is the secondary inference looking for some kind frame_user_meta, or obj_user_meta data?

[abdo.babukr1]

Here is what we are basing our python application from, how could this be used with a secondary network?

deepstream_python_apps/deepstream_ssd_parser.py at master · NVIDIA-AI-IOT/deepstream_python_apps (github.com)

[Fiona.Chen]

@abdo.babukr1

So my question is whats missing here?

Can you provide your code and configurations? We don’t know what kind of DsUserMeta you expected. Your description is confusing.

@abdo.babukr1

Is the secondary inference looking for some kind frame_user_meta, or obj_user_meta data?

No. SGIE works only on the objects detected by PGIE.

[abdo.babukr1]

this needed to be done in python for secondary inference to work:
obj_meta->unique_component_id = meta->unique_id;