Skip to main content

Accuracy Improvement

Accuracy Improvement turns the validation work you already do into new reference images for your catalogue. Every correction you make while reviewing results is a human-verified example of what a product actually looks like on shelf — which is exactly the data the recognition models need most.

How it works​

  1. You validate results. Review the predictions on a result and correct anything wrong.
  2. ZIA captures the corrections. Each validated annotation becomes a candidate reference crop.
  3. Similar crops are clustered. Candidates for the same product are grouped together.
  4. Clusters become submissions. Each group is sent to the accuracy improvement channel in the Catalogue Gatekeeper.
  5. You review and approve. Approved crops are added to the relevant item or variation.
  6. Accuracy improves. Later images are matched against a richer, more representative reference set.
tip

Validation does double duty — it corrects the result you are looking at and produces training data that improves every result after it. Corrections on products that the models currently get wrong are worth far more than confirmations of products they already get right.

What counts as validated data​

A crop becomes a candidate when you actively validate and save. That includes:

  • Correcting a misidentified product
  • Adding a detection the localiser missed, by drawing a bounding box
  • Adjusting a bounding box that was inaccurate
  • Explicitly confirming a prediction

Simply opening a result and looking at it does not generate anything — the change has to be saved.

note

Corrections are the most valuable signal, because they mark the cases where the models were wrong. A workflow that only confirms correct predictions will produce far fewer useful reference images.

Preparing items for review​

Validated crops are not sent for review automatically. You choose when to gather them and which ones to include, from the Validation Results widget on the Task Management page.

The widget shows:

  • IR Results Validated — how many results you have validated
  • Reference Images Detected — how many candidate crops that produced
  • Items Identified — how many distinct products appear in those crops

Click Prepare items for review to open the filter dialog.

Filters​

Time period of validations — the date range of validations to include. Recent validations reflect your current catalogue and current shelf reality, so a rolling recent window is usually a better starting point than everything ever validated.

Tasks to include — all tasks, or a specific selection. Narrowing to particular tasks is useful when you want references from one store format, region or product set rather than a blend of all of them.

Validation type — restrict to annotations that were user corrected only, excluding straightforward confirmations. This concentrates the batch on the cases where the models struggled.

Once prepared, the resulting clusters appear in the Gatekeeper for approval.

Getting the most from it​

  • Validate consistently rather than exhaustively. A steady trickle of corrections across many stores and conditions beats one large burst from a single visit.
  • Prioritise your problem products. If a handful of items drive most of your errors, validating images containing those items is the fastest route to improvement.
  • Be selective at approval. The reference image cap means each slot counts — favour crops that add a new angle, light level or shelf context over ones that resemble existing references.
  • Expect a lag. Improvements apply to images processed after the references are approved. To apply an improved catalogue to images you have already sent, see Reclassify Old Results.