Sgie’s detecting properly but classifier_meta_list is None

[monitarb]

Hey, first of all, thanks for doing this in Python!

I'm trying to replicate this in my Jetson NX with a couple of changes:

• My pgie model is the 4-classes model used on python-apps
• My sgie model detects mask/not_mask so it operates on people (class "person")
• I do not need a ROI in my case nor the OCR model for LP Text extraction... So I made the changes needed on the code and to my config files. I can see the mask/not mask Bounding Boxes in my display, but I can't get ahold of the metadata inside the code. This is: obj_meta.classifier_meta_list is always None. I have changed the inside of the if condition to check this:

l_class = obj_meta.classifier_meta_list
while l_class is not None:
	print("There's SOME metadata!")

And I can see the sgies bounding boxes on the display, inside some person's bounding boxes, but I NEVER see my "There's SOME metadata!" message. I did this test because I need the metadata of my sgie inside l_class. Have you come across this? What could be the issue?

[preronamajumder]

Hi, there is some issue regarding fetching classifier metadata that I have observed. l_class gives data for this LPR and for the caffemodels (the ones in test-app2 regarding car color, make and model). But for some reason it gives None for other classifiers. There are some etlt classifiers in NVIDIA NGC which I tried as secondary but they also gave l_class = None. I havent had the chance to explore it more or raise the issue in the forum. Maybe you should try posting the question in the forum.

[monitarb]

Hey there! I think I made some progress on my side, but now I've come to another block I hope you could help me: In my case classifier_meta_list is None because my sgie is not a classifier. My pipeline is: pgie (detector:4 classes) -> tracker -> sgie (detector on class Person: Mask, helmet, boots) It would like if, in your example, your sgie1 looked for licence plate, lights and wheels on the class Car... right? I believe I'm looking for the metadata of my sgie in the wrong place since classifier_meta_list gives the metadata of the classifier (In your case, the letters of the licence plate). Now, since I don't have a classifier, any idea where to get the metadata from the sgie1 (detector) from the probe?

[preronamajumder]

If you put the following you can single out data from individual gie. for primary you should do the following:

if obj_meta.unique_component_id == 1:
        class_id = obj_meta.class_id
        x = obj_meta.rect_params.left
        y = obj_meta.rect_params.top
        w = obj_meta.rect_params.width
        h = obj_meta.rect_params.height

for sgie it should be the following:

if obj_meta.unique_component_id == 2:
         class_id = obj_meta.class_id
         x = obj_meta.rect_params.left
         y = obj_meta.rect_params.top
         w = obj_meta.rect_params.width
         h = obj_meta.rect_params.height

the unique_component_id corresponds to the gie-unique-id provided in the primary/secondary config

This is a link to the python APIs for NvDsObjectMeta. You can see here what metadata you can extract. https://docs.nvidia.com/metropolis/deepstream/python-api/PYTHON_API/NvDsMeta/NvDsObjectMeta.html

[monitarb]

Thank You! This is actually what I was looking for! FYI, obj_meta gives the correct labels (either obj_meta.unique_component_id == 1 or 2) on my detector model trained on TLT! In my case (my sgie is a detector), classifier_meta_list is always None and this was Ok!

[vinodbukya6]

Hi, thank you for this project. I am facing the same problem as above l_class = obj_meta.classifier_meta_list l_class is always None. Can you please help me on this?

[monitarb]

In my case it was because my sgie was a detector model, NOT a classifier. So classifier_meta_list for it always returned None. In that case, the metadata I was looking for could be found in object_meta as well. The way to distinguish it from the metadata of pgie was using obj_meta.unique_component_id == 1 (pgie) or 2 (sgie) as @preronamajumder shows above. Good luck!

[vinodbukya6]

Thank you. Yeah, in my case it should work because it is a classifier(car license plate recognization). I tried with NVIDIA reference also (https://github.com/NVIDIA-AI-IOT/deepstream_lpr_app) still facing the same issue.