Import model predictions as staged annotations

Convert an integer prediction mask into ROIs and stage them against a project; nothing is committed. The prediction must match the image dimensions at level, contain only integers, and every non-background value must be in label_map; otherwise the import aborts. Optional per-pixel confidence below min_confidence is set to background first. Model and prediction digests and the confidence policy are recorded on every ROI, which starts as review_status = "unreviewed". Reviewed or locked annotations are never overwritten: with replace = TRUE they become staged conflicts.

Usage

at_training_import(
  x,
  predictions,
  label_map,
  level = 0L,
  source = "dnn_prediction",
  model = list(name = NA_character_, version = NA_character_, sha256 = NA_character_),
  confidence = NULL,
  min_confidence = NULL,
  layer = "predictions",
  replace = FALSE,
  stage = TRUE,
  call = rlang::caller_env()
)

Arguments

x

An annot_project.

predictions

A mask file (TIFF/PNG/NPY), an annot_mask, or an integer matrix [y, x].

label_map

Named integer vector label = code (codes >= 1; 0 is background).

level

Pyramid level of the prediction grid.

source

ROI source string. Default "dnn_prediction".

model

List with name, version and sha256 of the model.

confidence

Optional numeric matrix of per-pixel confidence.

min_confidence

Optional threshold applied to confidence.

layer

Target layer name. Default "predictions".

replace

Logical; stage removal of existing unreviewed ROIs of layer not present in the prediction.

stage

Logical; return an at_staged_patch (default) or, with FALSE, the prediction layer only.

call

The calling environment, for error reporting.

Value

An at_staged_patch (or an annot_layer when stage = FALSE) with attribute prediction_digest.

See also

at_commit_qupflowr()

Other training: at_training_check(), at_training_export()