DeepStream PPE Detection - Same Class ID for Person and Goggles Causing Conflict

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Hi @junshengy,

I’m currently developing a PPE detection application using DeepStream 6.3 on Jetson Orion, and I’m facing an issue when trying to differentiate between detections of “person” and “goggles” from two different models. Both models assign class_id == 0 to their primary objects.

Pipeline Overview:

  • PGIE (Primary Inference - Person Detection)

    • Custom YOLO model

    • Detects only person

    • class_id == 0, gie-unique-id = 1

  • SGIE (Secondary Inference - Goggles Detection)

    • Custom YOLO model trained on internal dataset

    • Detects goggles, shoes, etc.

    • Goggles also assigned class_id == 0, gie-unique-id = 2

The Issue:

Initially, I used the following logic in my probe to identify both person and goggles:

in pgie1_probe

if obj_meta.class_id == 0:
    person = True

in sgie1_probe

if obj_meta.class_id == 0:
    
    goggles = True  # <-- Conflict happens here

As expected, this caused incorrect results — when a person was detected, it also showed goggles as detected, even if the person wasn’t wearing any.

Attempted Fix:

To resolve the conflict, I tried using unique_component_id (which should match the gie-unique-id from the config):

My gie-unique-id for Human model is 1 and goggles model is 2

# For person detection in pgie1_probe
if obj_meta.class_id == 0:
    person = True

# For goggles detection in sgie1_probe
if obj_meta.class_id == 0 and obj_meta.unique_component_id == 2:
    goggles = True

Now the goggles detection always returns false, even though I can visually see the bounding boxes for goggles on the output screen

My Questions:

  1. Is using obj_meta.unique_component_id the correct way to differentiate between detections from different inference engines (PGIE vs SGIE), especially when they share the same class_id?

  2. Is there a better way to handle this situation where two different models use the same class_id?

Any insights or suggestions would be greatly appreciated.

Thanks,

Hema.

2

Yes. You can use unique_component_id to identify which model detect the object.

You have done the correct thing.

4

Hi @Fiona.Chen ,

Thanks for the reply.

When I use unique_component_id to differentiate detections, as discussed above:

if obj_meta.class_id == 0 and obj_meta.unique_component_id == 2:
    goggles = True

However, the condition for goggles is always evaluating to False. I also tried printing the unique_component_id like this:

model_unique_id = obj_meta.unique_component_id
print("Model Unique ID:", model_unique_id)

But nothing gets printed — it’s as if unique_component_id isn’t being set in the object metadata.

From what I’ve researched, it seems this might happen if the object metadata isn’t being attached correctly, or if the inference component isn’t populating unique_component_id. That said, the model itself is working fine and detecting objects as expected.

Can you suggest what steps I can take to debug this? Specifically:

  • How can I confirm that the metadata is being added correctly?

  • Is there a specific configuration required to enable unique_component_id propagation from a secondary model?

  • Any tips on verifying whether the inference component is actually setting this field?

I already tried cross checking the class id of goggles which is 0 and goggles model config file where gie_unique_id is 2 (which should be equal to the unique_component_id).

5

The object meta is attached by nvinfer plugin, the plugin (/opt/nvidia/deepstream/deepstream/sources/gst-plugins/gst-nvinfer) and its library(/opt/nvidia/deepstream/deepstream/sources/libs/nvdsinfer) are all open source, you can debug it directly.

6

Thank you once again.

Based on my research, it looks like there are other metadata components available, such as parent, which can help with differentiating objects when using multiple models.

In my case, the goggles detection model is operating on the human detection model as a secondary classifier. Here’s how it’s defined in the goggles model config:

operate-on-class-id=0
operate-on-gie-id=1

From what I understand is :

can I use something like this to differentiate goggles detections?

if obj_meta.parent == 1 and obj_meta.class_id == 0:
    # This should correspond to goggles on a human

Would checking obj_meta.parent in this way be a reliable approach to distinguish between detections from different models when unique_component_id is not available or not working as expected?

7

I don’t understand your design. Seems you have two detection models, what is the relationship between the two models in your application?

8

You’re correct — I have two models in my application.

  • The first model is a Primary GIE (PGIE), which performs detection on the full frame and identifies humans.

  • The second model is a Secondary GIE (SGIE), which runs on the bounding boxes output by the PGIE. This SGIE detects goggles and shoes specifically on the detected humans.

So, the SGIE depends on the results from the PGIE — it operates only on objects detected as humans.

9

There is similar PGIE+SGIE sample in deepstream_lpr_app/deepstream-lpr-app/deepstream_lpr_app.c at master · NVIDIA-AI-IOT/deepstream_lpr_app, we can get the “unique_component_id” in the probe function.