Post-Processing for DeepStream App

Accelerated Computing Intelligent Video Analytics DeepStream SDK
, , , ,
1

Hi,

I am using DeepStream App for one of my AI Object Detection Application. I am running this on Jetson Orin Nano-8 GB kit with Jetpack 6.1. I am using Ultralytics Yolov11 trained on our custom dataset for the same. I have found the steps from Ultralytics official documentation to run YOLOv11 with DeepStream. I have followed same repo to do that. I have attached the link of that github repo below.

I am getting proper output with this. The entire pipline is running properly. This is the results I am getting.

deepstream_yolo
deepstream_yolo973×863 42.6 KB

Now, I want to modify post-processing steps further to make the bbox and label with diff. colors as well as I want to add object tracking as well. I want to count the objects present and want to display in real-time in table format. How to do that with this pipeline? From where we can modify the post-processing steps ?

I have done the same thing on other platform, I am attaching the snapshot of that for reference.

gst-pill-detection
gst-pill-detection758×566 70.9 KB
3

You can try to add a tracker-group and a nvds-analytics-group in the deepstream_app_config.txt. The analytics part is open-source and you might need to customize it according to your specific scenario.

4

Hi @yuweiw ,

Thanks for the quick response. In my case, I need the stats like how many object of perticular class is present on screen irrespective of unique id to each object. I think you can get the point by looking at second image I have attached in previous chat.

Can you please guide how to create this type of table in overlay and how can we add multiple colors in bounding box? Which pluging or config I need to modify for that? And how? If it’s open-source then where to get source for the same?

Please let me know if you have still confusion.

5

Since you don’t want to get the unique id, there is no need to use a tracker in this scenario.

The nvdsosd can meet your needs. You can just refer to our deepstream\sources\apps\sample_apps\deepstream-test1\deepstream_test1_app.c sample to learn how to get the number of the object and draw the NvOSD_TextParams.

static GstPadProbeReturn osd_sink_pad_buffer_probe (GstPad * pad, GstPadProbeInfo * info,     gpointer u_data)
6

Hi @yuweiw ,

Thanks for the reply. Let me try that.