deepstream-dev

Use the pyservicemaker Pipeline API to add elements like nvurisrcbin, nvstreammux, nvinfer, and a display sink, then link them and call start(). Sources should default to nvurisrcbin, which handles RTSP, HTTP, and local files transparently.

Point the nvinfer config at your ONNX file and set infer-dims for dynamic models, such as infer-dims=3;640;640. Match cluster-mode to the model generation: v8/v11 output raw pre-NMS tensors needing cluster-mode 2, while v10+ output post-NMS results needing cluster-mode 4.

Pipelines using tee splits or dynamic sources hang in PAUSED when any sink lacks async=0. Set async: 0 on every sink element, including nvmsgbroker and display sinks, to allow the state transition to PLAYING.

pyservicemaker is installed system-wide but is not visible inside a standard venv. Install it explicitly with pip install /opt/nvidia/deepstream/deepstream/service-maker/python/pyservicemaker*.whl pyyaml inside the venv.

Yes, use nvmsgconv to convert metadata and nvmsgbroker with the libnvds_kafka_proto.so protocol library to publish to Kafka. Since nvmsgbroker is a sink, use a tee element to split the pipeline when you also need video output.