Deepstream_test1 obj dection two level threshold

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

Please provide complete information as applicable to your setup.

• Hardware Platform (Jetson / GPU) Agx Xavier
• DeepStream Version 5.0
• JetPack Version (valid for Jetson only) 4.4
**•

I am running deepstream_test1 and using the default dstest1_gie_config.txt configure file. It runs well. But I notice the car or person detection are a little bit “jumpy”: that is the bounding box sometime disappear and reappear again on a same object for a given video. I understand that’s “normal” given the particular model training sensitivity/specificity. However, I wonder if there is a parameter setting I can set in the configure file to have two level of object detection thresholds to “smooth out” this jumpy behavior: let’s say hi-threshold=0.5 and lo-threshold=0.35. When the object detected with probability >= hi-threshold, the object is considered detected or if the current object probability < hi-threshold but > lo-threshold and the previous frame same object probability > hi-threshold, then the current object is still considered detected. Could this be done?

Or this require to modify the model logic a little bit and how? Could you advise the resource that I can learn from to accomplish this?

Thanks a lot for your help.

3

Currently we don’t have such configs.
I think you can change your pipeline and add the tracker following the detector, you can use nvdcf tracker and also set the nvinfer’s property interval > 0 and see if the bbox more soothy.