How to use PeopleSemSegNet model after using normal object detection model

Accelerated Computing Intelligent Video Analytics DeepStream SDK
#1

• Hardware Platform (Jetson / GPU) Jetson AGX Orin
• DeepStream Version 6.1
• JetPack Version (valid for Jetson only) Unknown
• TensorRT Version 8.4
• Issue Type( questions, new requirements, bugs) questions

Hi, I’m trying to mask people in a video and now I’m having issues with its accuracy. Model which I use (peoplesemsegnet) doesn’t mask all the people.

However, some object detection models can detect all the people in the image and draw valid bounding boxes.
So I want to use these object detection models first, and use PeopleSemSegNet after the detection. That is, I want to specify areas in which a person is standing so that PeopleSemSegNet is more likely to mask people correctly.

Therefore I have this question: Is it possible to specify ROIs when using PeopleSemSegNet with nvinfer?

Give me some hints. Best regards.

#3

Please use PeopleSemSegNet as secondary GIE. Gst-nvinfer — DeepStream 6.1 Release documentation

We already have a lot of SGIE samples. E.G. deepstream-test2
https://docs.nvidia.com/metropolis/deepstream/dev-guide/text/DS_C_Sample_Apps.html

#4

Thanks. I’ll try it and reply results then.

#5

I referred deepstream-test2 in deepstream_python_apps because I use Python Bindings. And I was able to use SSD(as PGIE) and PeopleSemSegNet(as SGIE) in the pipeline. (I added nvtracker, too.)

But I don’t know how to set ROIs in secondary GIE. Could you give me additional resources which help me to achieve my goal?

#6

Is the ROI fixed?

#7

No.
I have completely no idea how to set ROIs or cutomize settings related to ROIs.

#8

It seems nvdspreprocess plugin satisfies my requirements. I’ll try the plugin.

#9

ROIs can not be set dynamically.

#10

Oh, I didn’t know that. Maybe, I have to implement the feature by myself…