Sgie inference does not work for some detected objects

Please provide complete information as applicable to your setup.

**• Hardware Platform: Jetson nano
**• DeepStream Version: 6.0
**• JetPack Version
• TensorRT Version
• NVIDIA GPU Driver Version (valid for GPU only)
• Issue Type( questions, new requirements, bugs)
• How to reproduce the issue ? (This is for bugs. Including which sample app is using, the configuration files content, the command line used and other details for reproducing)
[dstest2_pgie_config.txt]
pgie.txt (1.2 KB)

[dstest2_sgie1_config.txt]
sgie.txt (3.2 KB)

• Requirement details( This is for new requirement. Including the module name-for which plugin or for which sample application, the function description)

I used nvidia peoplenet for detecting full-length body and face and developed a gender classifier as the secondary network. When I set the operate-on-class-ids=1 to pass the face bounding box to the gender classifier nothing happens and there is no gender associated with the face. When the value is set to 0 (person is passed over to the classifier) then I can see the gender label appearing next to the person object (although this is not the proper use as I trained the gender classifier on faces only). Can you please help with this issue.

I fixed the issue as 1 refers to bag class and not face, but now the problem is that the classifier is always predicting Male gender for all the persons. Any suggestion?

We don’t know anything about your model. Maybe the model precision debug tips is helpful for you. DeepStream SDK FAQ - Intelligent Video Analytics / DeepStream SDK - NVIDIA Developer Forums

I developed my model on TAO and the validation results are pretty high and I created te etlt file using the TAO exporter. Any idea how can I control the size of the image that is passed through to the sgie inference?

Do you know the input layer dimension of your model? gst-nvinfer will do the correct preprocessing if you configure the “infer-dims”, “model-color-format”, “net-scale-factor” and “offsets” correctly.
https://docs.nvidia.com/metropolis/deepstream/dev-guide/text/DS_plugin_gst-nvinfer.html

yes, it’s 3;224;224

gst-nvinfer can do format converion, normalization and scaling preprocessing.

I input all of these parameters and I suppose they should be correct as I got them from the file generated at the export time

We don’t know anything about your model. Maybe the model precision debug tips is helpful for you. DeepStream SDK FAQ - Intelligent Video Analytics / DeepStream SDK - NVIDIA Developer Forums

As I already explained, I entered all the required parameters as requested but still nothing happened. I tested with another video (with larger faces) and the classification doesn’t seem to work at all for this case. Can you please have a look at my config files (I attached them to my post) and tell what should I change if any

can I pass you the model file to reproduce the case and see where it goes wrong

There is no update from you for a period, assuming this is not an issue anymore.
Hence we are closing this topic. If need further support, please open a new one.
Thanks

All required parameters depend on your model.

All parameters definition can be found in Gst-nvinfer — DeepStream 6.1.1 Release documentation

Please do not use “mean-file” and “offsets” at the same time. There is no calibration file, so the “network-mode” can not be INT8.

You set “operate-on-class-ids=1” in sgie.txt and your PGIE is peoplenet, we don’t know whether it is the sample peoplenet from TAO(PeopleNet | NVIDIA NGC), if it is, the peoplenet class id 1 means “bag”, do you want to classify the “bag”?

For other parameters, since we don’t know about your model, please check with the guy who provide the model to you.

Please read the document carefully. And we already provide samples for PGIE and SGIE. The gst-nvinfer is based on tensorRT and totally open source. You can debug with the source code and DeepStream SDK FAQ - Intelligent Video Analytics / DeepStream SDK - NVIDIA Developer Forums may also help.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.