Skip to main content

Reclassify Old Results

Improvements to your catalogue normally only affect images processed after the change. Reclassification closes that gap: it re-runs classification on results you have already received, against your catalogue as it stands today.

This is the natural companion to catalogue growth. If Real Data Flow has just onboarded two hundred competitor products, every image you sent last quarter contains those products labelled as unknown. Reclassification lets you recover that history instead of writing it off.

warning

Reclassification is currently a back-office feature and is triggered by Neurolabs on request — there is no self-serve control in the web app. Some parts of the workflow are still being completed (see Current limitations). Treat it as available on request rather than generally available.

When it is worth running​

  • A significant catalogue expansion has happened, and historical results are full of unknowns that are now recognisable
  • Reference images have been added to products that previously performed poorly
  • Products have been rebranded, and older results were matched against superseded artwork
  • You want to quantify what a catalogue improvement was actually worth

How it works​

Reclassification reuses the visual signatures (embeddings) computed when your images were first processed. Nothing is re-detected and no image is reprocessed — only the matching step runs again, against your current reference set.

That has two useful consequences: it is far cheaper and faster than resubmitting the images, and it isolates the change. Because detection is untouched, any difference in the output is attributable to the catalogue, not to a different set of bounding boxes.

The typical sequence is:

  1. You send images over a period of time, and receive results.
  2. Your catalogue is improved — new items, new reference images, corrected metadata.
  3. Neurolabs triggers reclassification for your organisation over a chosen date range.
  4. New results are created carrying the reclassified predictions.
note

Because detection is not re-run, reclassification cannot recover products the localiser never detected in the first place. If a product was missed entirely — not detected, rather than detected and mislabelled — reclassification will not find it, and the image needs to be reprocessed properly.

The rematch report​

Before anything is written, a rematch report can be produced as a dry run. It re-scores your previously validated annotations against your current reference set and classifies each one, without changing any result.

Every annotation is labelled with one of:

OutcomeMeaning
FIXEDWas wrong before, correct now — the catalogue change helped
REGRESSEDWas correct before, wrong now — the catalogue change hurt
STILL_CORRECTCorrect before and after
STILL_WRONGWrong before and after — not addressed by this change
NO_EMBEDDINGNo cached signature available, so it cannot be re-scored
MISSED_BY_LOCALISERNever detected in the first place, so out of scope

The report is delivered as a CSV covering an organisation and date range.

tip

Run the rematch report first, and read the REGRESSED count before committing to a reclassification. A catalogue change that fixes 400 annotations and regresses 300 is not the win the headline number suggests — and regressions usually point at a specific bad reference image that is worth removing before you proceed.

Current limitations​

Reclassification is functional but not finished. Known gaps:

  • Mapping old results to new. Reclassification creates fresh results rather than updating existing ones. Reliable linking between an original result and its reclassified counterpart is still being completed, so reconciling the two in your own systems needs care.
  • Callbacks. Result callbacks are not yet dependably fired for reclassified results — do not assume your webhook will notify you.
  • Reporting. A dashboard for reviewing reclassification outcomes is in progress. Today, the rematch report CSV is the primary output.
info

Given these gaps, agree up front with your Neurolabs contact how the reclassified results will be delivered and how you will reconcile them with the originals — particularly if your integration reads results via callbacks.