Custom DeepStream Deep Learning App

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Hi,

I have a Nvidia Jetson Nano and I want to create an application, preferably in Python, to track and re-identificate people in multicamera scenarios. Following deepstream-test3.py I am able to connect 4 RTSP cameras, detect people, by using resnet caffe model, and draw their bounding boxes. The problem and my questions come when I want to:

  1. Crop the people when I click over them in the frame in order to apply another CNN. I tried with gst_navigation_event_parse_mouse_button_event but it does not work. Is this possible in GStreamer and DeepStream?

  2. Convert my siamese neural network, developed in Tensorflow, to a deepstream plugin. How can I migrate the tensorflow CNN to deep stream, it is necessary to developed in C++?

2

Hi,

You may refer to test2 sample, which has one primary model(pgie) and three secondary models(sgie). You can make ‘another CNN’ into a secondary model.

We have a reference of converting TensorFlow to TRT. Please check

/usr/src/tensorrt/samples/sampleUffSSD/
3

Hi,

The problem is that I need to apply the sgie (my custom sgie) just to some bbox, not all of them, so:

  1. How can I select the bboxes I want as input in the sgie?

  2. How can I make a GUi with Ds and Osd in order to select with a click the person to track? There is any library for that?

On the other hand, I will see the reference that you told me about converting .pb to .uff.

Regards,

4

Hi,
You would need to implement interactive GUI. Not sure if there is existing implementation in python. Other users may share experiences/guidance on this function.

For selecting bboxes, you can get all bboxes and then remove those you don’t need. There is a sample of using nvds_remove_obj_meta_from_frame():
https://devtalk.nvidia.com/default/topic/1070986/deepstream-sdk/function-nvds_remove_obj_meta_from_frame-in-python-/post/5426343/#5426343
Please look at the post and give it a try.

5

Ok, it works.

Thanks.