streamvision

streamvision holds the host-neutral pieces of Roboflow Inference's video stack: camera acquisition (VideoSource and its producers), the host-neutral InferencePipeline, the stream-manager TCP client and its wire entities, and the stream-manager runtime that hosts pipelines as subprocesses. It is a sibling distribution to roboflow-workflows, installable on its own or embedded in the full Inference server.

The historical inference.core.interfaces.{camera,stream,stream_manager} import paths resolve to the same modules here when inference is installed.

Configuration

Settings live in one StreamsConfiguration per process. Install it with streamvision.stream.configuration.configure_process before importing any runtime module: the values are frozen into module constants at the first import. Without an installed configuration the defaults apply. Installing a different configuration afterwards raises StreamsConfigurationError, which names the differing fields.

In a process that also uses inference, import it first:

import inference.core  # or: from inference import InferencePipeline
import streamvision.stream.pipeline

inference.core installs the configuration built from the environment variables of inference. A plain import inference is lazy and installs nothing. With the opposite order and an environment that differs from the defaults, import inference.core raises StreamsConfigurationError or WorkflowEnvironmentConfigurationError.

Processes started by the stream manager install the configuration passed by the launcher before they import the host factory module.

Hardware decoding

Hardware decoders are optional. Without them VideoSource decodes on the CPU with OpenCV. They are tried only when enable_tensor_data_representation is set in the installed StreamsConfiguration. A decoder that cannot be used is skipped with a warning that names the reason.

Standalone stream manager

The stream manager (python -m streamvision) needs streamvision[webrtc,workflows]; the library parts (streamvision.camera, streamvision.stream, the TCP client and entities) work without both extras; workflow pipelines need workflows.

STREAM_MANAGER_PORT=7070 python -m streamvision \
    --host-factory my_package.host:create_host \
    --host-setting api_key=... --warm-pipelines 1

--host-factory names a trusted callable returning a pipeline host; its module is imported before the runtime. scripts/streamvision_isolation_probe.py verifies an installed wheel runs without inference.

Installation extras

Without workflows, cameras, InferencePipeline.init_with_custom_logic, sinks, the watchdog and the stream manager client work. The two workflow functions raise CannotInitialiseModelError with an install hint, and python -m streamvision exits with a message naming the missing extra.

Without workflows, concurrent.futures.Future objects returned by custom logic reach the sinks unresolved; with it they are resolved as before.