ml-pipes-supervision 0.1.3

Roboflow Supervision capabilities as composable operators in ml-pipes.
Build explicit computer-vision pipelines for detection, annotation, tracking, and zone counting.

Coverage

See coverage for role definitions and the detailed API compatibility matrix.

Install

Requires Python 3.10 or newer.

python -m pip install ml-pipes-supervision

For RoboflowInference, including the Quickstart below, install the optional Inference integration:

python -m pip install "ml-pipes-supervision[inference]"

Quickstart

Build the usual detection-and-annotation flow as one pipeline. The operators below are the same thin boundaries used by the runnable examples.

from ml_pipes.core import Pipeline
from ml_pipes.standard import Recall, Select, Store
from ml_pipes.vision import Decode, LoadFile
from ml_pipes.supervision import BoxAnnotator, Detections, ImageToArray, LabelAnnotator, PlotImage
from ml_pipes.supervision.inference import RoboflowInference

pipeline = Pipeline(
    [
        LoadFile(),
        Decode(),
        ImageToArray(),
        Store("source_image"),
        RoboflowInference(model_id="rfdetr-small"),
        Select(0),
        Detections.FromInference(),
        Recall("source_image", prepend=True),
        BoxAnnotator(),
        LabelAnnotator(),
        PlotImage(),
    ]
)

For model integrations that produce an ml_pipes.tensor.TensorRegistry, use Detections.FromTensorRegistry() before Supervision annotators, trackers, zones, or sinks. This is the explicit boundary from tensor post-processing to Supervision data.

Documentation

Want to learn how to use Supervision with ml-pipes? Explore our how-to guides and end-to-end examples!

The GitHub Pages tutorials preserve the corresponding Supervision guides and add their ml-pipes counterparts, so you can compare both approaches side by side.

Why Supervision with ml-pipes?

Supervision provides the computer-vision building blocks; ml-pipes makes the boundaries between those blocks explicit, composable, and inspectable. A pipeline can validate its contracts before execution and capture the value at each operator boundary with Pipeline.inspect(). That makes complex flows easier to understand and debug without adding ad-hoc logging to every step.

The Detect Small Objects pipeline is a good example: it tiles the input image, runs inference on each tile, gathers the results, stitches detections back into the source coordinate system, merges overlaps, and annotates the final image. The inspection report shows every boundary in that flow.

Click the image to open the interactive inspection report.

Built with Supervision x ml-pipes