Skip to main content

Review (Catalogue Gatekeeper)

The Catalogue Gatekeeper is the review step that every piece of incoming catalogue data passes through before it reaches your live catalogue. Nothing is added silently: proposed reference images, new items and metadata updates all arrive as submissions for a human to approve or reject.

This matters because reference images are what the recognition models match against. A single mislabelled crop approved into an item degrades accuracy for every image processed afterwards — so the Gatekeeper is deliberately the one place where catalogue growth is gated.

Submissions​

A submission is a proposed change to your catalogue, bundled for review. Depending on where it came from, it may contain:

  • Candidate reference images for an existing item or variation
  • A proposed new catalogue item, with clustered crops and suggested metadata
  • Updated item properties
  • A 3D asset or video capture awaiting processing

Each submission can be approved (the content is written into the catalogue) or rejected (it is discarded and does not affect recognition).

Channels​

Submissions are organised into channels by where they came from. Each source has its own channel so that reviewers can apply the right level of scrutiny to each, and so that a noisy source can be paused without blocking the others.

ChannelWhat arrives here
Real Data FlowClustered crops from your own shelf images — new item candidates and extra references for existing items. See Real Data Flow
Accuracy improvementCrops derived from results you have validated. See Accuracy Improvement
Synthetic dataSynthetic reference images generated from artwork or 3D assets. See Synthetic Data
ZIA CaptureProduct captures and 360° video submitted from the field
Item formProposed item and metadata updates submitted directly

An organisation has one channel per source. Channels can be configured so that submissions require moderation, or so that they are approved automatically on arrival.

warning

Auto-approval is applied when a submission is created. Switching a channel to auto-approve does not retroactively approve submissions that are already sitting in the queue — those still need to be actioned, in bulk if necessary.

Reviewing submissions​

  1. Open the Review area and select the channel you want to work through.
  2. Open a submission to see the proposed images and the item or variation they would be attached to.
  3. Check each crop before approving:
    • Is it definitely the right product, including size and flavour variant?
    • Is the crop clean — the product in full, without neighbouring products included?
    • Is it different enough from existing references to add value?
  4. Approve or reject.
tip

Rejecting a weak crop costs nothing. Approving one is hard to undo, because it immediately becomes part of what the models match against. When a crop is ambiguous, reject it and wait for a better one.

Fast Review​

For high-volume channels, the Fast Review tab supports bulk approval and rejection so that a large batch of clearly-good or clearly-bad submissions can be cleared without opening each one individually. Use it for triage, not as a substitute for inspecting anything you are unsure about.

Reference image limits​

Each variation accepts a maximum of 25 reference images. Approval is blocked when the images in a submission would push a variation past that cap, with an error asking the reviewer to submit fewer.

This is a quality constraint rather than a storage one. Beyond roughly this number, additional near-duplicate references stop improving recognition and start crowding out the diversity that actually helps — a spread of angles, lighting conditions and shelf contexts beats a large pile of similar crops.

note

If a variation is already at the cap and you have a better reference image, remove a weaker existing reference first, then approve the new one.