TaughtHuman decisions for better models

Image annotation that stays connected to the model.

Taught gives object-detection teams one traceable path from source images and machine proposals to human-reviewed boxes, frozen datasets, and promoted model versions.

Request early access Explore open models
Evidence registerA box is more useful when its history survives.

The review record distinguishes what the model suggested, what a person changed, and what the project accepted.

  1. Source
    Original image

    Identity, dimensions, grouping, and provenance remain attached.

  2. Proposal
    Model suggestion

    Candidate class, geometry, score, and checkpoint stay separate from human evidence.

  3. Review
    Human decision

    Confirm, reject, adjust, add, or mark the frame exhaustively negative.

  4. Register
    Eligible annotation

    Task rules and review gates decide what may enter a frozen dataset.

Built for object-detection evidence, not anonymous rectangles.

Annotation quality depends on more than drawing controls. Taught preserves the task, taxonomy, source, reviewer action, and eligibility decision around each box so teams can reproduce how training data was formed.

Normalized geometry
Boxes retain a defined coordinate space so exports and serving results can be compared consistently.
Exhaustive review
A reviewer can explicitly record that no eligible object was found instead of leaving an ambiguous empty frame.
Proposal isolation
Machine suggestions accelerate review without becoming votes or silently entering ground truth.
Source grouping
Acquisition groups support leakage-aware dataset splits instead of randomizing near-duplicate frames across train and test.

From raw frame to reproducible detection dataset

  1. 01
    Define the detection task

    Freeze the taxonomy, geometry rules, review instructions, and eligibility gate before collecting evidence.

  2. 02
    Review proposals independently

    People judge the source without inheriting another reviewer’s vote direction.

  3. 03
    Resolve disagreement

    Conflicts remain visible for adjudication rather than disappearing into a majority label.

  4. 04
    Freeze and export

    Only eligible evidence enters a versioned dataset with split and provenance records.

What this page does—and does not—promise

  • Object-detection image review is a mature native workflow in Taught.
  • Model proposals remain suggestions; they never count as human votes.
  • Training and serving depend on the configured task, model provider, and deployment capabilities.

Questions teams ask before they begin

Direct answers based on the current Taught product and deployment boundaries.

Can reviewers add boxes the model missed?

Yes. A review can confirm, reject, adjust, or add an annotation under the project’s frozen task rules.

Can Taught record images with no target objects?

Yes. Exhaustive negatives are explicit evidence, not merely frames with an empty annotation list.

Does accepting a proposal make it a human vote?

The person’s recorded action is the human evidence. The proposal itself remains separately attributable to the model version that produced it.

Bring the decisions, data, and model version into one record.

Request access for a labeling project, a custom-model workflow, or an integration with a promoted model.

Request early access