I am currently working on a project where I need to create an application that can receive new configuration messages on the fly and update the DeepStream pipeline accordingly. Specifically, I would like to handle scenarios where the pipeline starts when a new stream comes in, but there’s a need to reconfigure the stream dynamically, for example, by changing the Region of Interest (ROI) size. Additionally, I am interested in dynamically removing streams from the pipeline, stopping analytics for the removed streams.
I have considered using Gst-nvdsanalytics, but it seems limited in terms of modification for new use cases beyond the ones provided in the documentation. In addition, is it possible to change the confidence threshold of the models while it’s running for one stream in a multiple stream app?
Are there any sample applications or documentation available that could provide guidance or inspiration for implementing dynamic configuration updates and stream management within a DeepStream application? I would appreciate any advice or resources you could share.
If you want to change ROIs with new stream, that means you have to update the context for the corresponding pad inside nvdsanalytics. What kind of issue have you met with modifying the nvdsanalytics source code?
The proper way is to set the pipeline to NULL state and you change the new source and set new configurations to any elements in the pipeline.
Do you mean you want to apply different cluster algorithms to different streams in the same batch? Currently it is not supported. You can implement by yourself by customize your own postprocessing with setting “output-tensor-meta=1” and “network-type=100” of gst-nvinfer. Then you can write your own postprocessing including cluster algorithm by yourself. Gst-nvinfer — DeepStream documentation 6.4 documentation
Thanks Fiona for the info I will take a better look on that.
No I don’t mean different algorithms, if I have the same classifier (nvinfer) running analytics for multiple streams can I change the confidence of the classifier for one stream and remain the same for the others.