Skip to main content

Detector and ingest policy

The backend can run a generic object detector when images are ingested, and a policy decides what an ingest keeps and which detections get an embedding vector. This page covers where you see that and change it. Every number, list and refusal shown is served by the backend; Cropwright computes none of them.

What each field means, the API calls and the cost of each mode are in the Ingest policy guide. Set the policy before a large ingest: it changes future ingests only.

The detector card (/ingest)​

The Ingest page shows a card for the served detector: model, version, input size, whether it assigns classes, and a collapsed table of its labels. It also gives a one-line summary of the current policy with a link to the policy page. With no detector reported, the card reads "No detector reported."

Ingest page, where the detector card and the policy link appear

The ingest policy page (/settings/ingest-policy)​

The form starts from the policy the backend serves.

  • Embedding mode, always open: all (every detection), selected (only those matching your criteria) or lazy (no detection vector until asked for; the whole-frame vector of each image is still computed) or none (detection only: no vector of any kind).
  • Clustering, a checkbox: "Cluster automatically". Unchecked, new items are not placed in clusters and the refresh and idle retrains skip the project.
  • Detect filter, collapsed under advanced: classes, excluded classes, confidence, area, a per-image cap and how classes are resolved.
  • Per-project detector override, also advanced.

Nothing is checked in the browser, so for example selected mode with no criterion is refused by the backend when you save.

Cost preview​

A moment after any edit the draft goes to the backend, which answers "N of M stored detections would be embedded, about X MB". The panel says outright that this describes detections already stored and that a policy change only affects future ingests. When the backend sampled, it says "estimated from S detections".

Ingest policy — selected mode, and the served cost preview: how many stored detections would be embedded and about how much space

Saving​

Save asks for confirmation, then writes the policy. If someone else saved first you are offered Reload (drops your edits) or Keep my edits (rebases them on the latest policy). A list of unknown class names is shown as a warning; the policy is saved anyway. A detector that cannot be served shows the backend's message and reasons.

Create classes from the detector (/classes)​

When a detector is reported, Classes has a collapsed panel that creates classes from the detector's labels. Pick labels (none means all), run Preview (a dry run), read which classes would be created, skipped or conflict, and only then confirm Create N classes.

See Embedding state and detections summary for what you can see afterwards, and how to fill in missing vectors.