Annotate: model-assisted labels and human review

Annotation turns an observation into a training target. For a perception model, that may mean boxes, masks, or tracked objects. For a robot policy, it may mean keypoints, actions, or contact phases. Start with the label definition and the model's needs.

Combine model assistance with human review

Supported auto-label models can produce initial predictions for compatible projects. Those predictions are a starting point: their quality depends on the model, the input data, and the task configuration.

Trained coworkers review labels against the project specification. They can correct predictions, resolve ambiguous scenes, and label work that is unsuitable for automation. Decide which outputs need review and how disagreements are resolved before the job starts. A confidence score is not a substitute for that policy.

Measure the usefulness of automation on representative data. Do not assume a fixed share of labels will be accepted automatically or that corrections retrain a model without a configured training process.

Check consistency across time and sensors

Physical AI tasks often require more than an independent label on each image. A stable object identity should refer to the same object across the frames in scope. A projected cuboid should agree with the sensor calibration. Interpolation should be checked where an object changes speed, disappears, or becomes occluded.

Avala's tools include 2D and 3D annotation capabilities. Choose the tools that fit the project and verify their outputs. Creating a label in one view does not guarantee correct propagation to every camera or timestamp, and annotation software does not eliminate calibration drift in the source data.

Define a review gate

Agree on acceptance criteria and a process for disputed labels. Where a project uses multiple annotators or consensus, inspect the disagreement rather than treating agreement alone as proof of correctness. Preserve the task and review history so the next team can understand what was accepted.

Start a supported auto-label job

Set AVALA_PROJECT_UID to an existing project UID in your workspace. Confirm that the project's data and labeling configuration support the selected model. This command creates a job and can write predicted annotations:

avala auto-label create \
  --project "${AVALA_PROJECT_UID:?Set AVALA_PROJECT_UID to your project UID}" \
  --model-type sam3

Use --dry-run to run inference without saving results. Inspect the resulting job and predictions before incorporating them into the reviewed training data. See the CLI reference for the available options.

Next: Curate & manage →