Data Labeling

navio provides an integrated data labeling workflow that lets you review and correct model predictions and use the resulting labels to continuously improve your models through retraining.

info

Data labeling is currently focused on image data and is used by navio VISION, navio's solution for visual anomaly inspection. For the user-facing labeling workflow, see Labeling & Retraining in the navio VISION documentation.

How It Works

The labeling integration is built around your use case: each use case with labeling enabled gets its own labeling project, powered by an embedded Label Studio instance.

1. A use case is created with a label configuration (currently image classification with the classes Anomaly and Normal). A labeling project is created for the use case automatically.
2. Connected image data is registered in the labeling project. Data connections can optionally carry a predefined label that is used as ground truth until a user label exists.
3. Model inference pre-labels the data. When a deployment runs inference over the datasets, a labeling task is created per image containing the model's prediction and confidence score — labelers confirm or correct suggestions instead of starting from scratch.
4. Labels flow back into navio automatically and are linked to the underlying datapoint and model prediction.
5. Retraining consumes the labeled data: navio assembles all labeled datapoints (user labels take precedence) into a training dataset, performs a train/validation split and retrains and redeploys the model.

Access to a labeling project is restricted to members of the corresponding workspace, using the navio login.

Current Scope

note

• Data labeling currently supports image classification (Anomaly/Normal).
• One labeling project is created per use case.
• The feature is optional and needs to be enabled for your instance — see Advanced Config. :::

If your use case requires other label types (e.g. object detection, segmentation, text or tabular data), get in touch with us.