#import "../../shared/macros.typ": *

=== Reliability and causal observation support

Confidence is useful here as an estimate of detection reliability, provided it
is calibrated for that task. Guo et al. study classification calibration and
temperature scaling @Calibration-guo2017; calibration of the present
three-dimensional detector remains to be established. Confidence stored with
a ground-truth box is a separate field and cannot substitute for the detector's
estimate.

Observed-to-ground-truth IoU describes localization overlap. It can associate a
detection with the correct evaluation object, but does not measure how much
its reconstruction could improve. Grouping training examples by ground-truth IoU is a
privileged data-selection procedure, even when the actor never receives IoU.
Such a procedure must leave the reference evaluation population unchanged and
cannot serve as an observation-only target-selection policy.

// evidence:
// - @Calibration-guo2017 -> https://proceedings.mlr.press/v70/guo17a/guo17a.pdf, Secs. 2 and 4 (classification calibration and temperature scaling).
// - aria_nbv/aria_nbv/oracle/target_selection.py: OracleTargetTask.confidence and match_observed_target_descriptors (payload confidence and privileged association).

Visible mask area describes how much of an image the target occupies. Adding
this area across frames measures accumulated exposure: repeated frontal views
increase the sum without revealing the cabinet's side. A projected OBB
footprint is less informative about visibility because it ignores occlusion.
Viewing-direction diversity and the union of observed surface elements offer
complementary descriptions of acquired evidence. These quantities could define
training strata, with repeated observations counted according to a declared
rule. They describe *observed support*, not full-object completeness unless the
total surface is known. Only evidence acquired by the current decision time may
contribute; missing masks or geometry remain unavailable measurements.

Jeong et al. use object probabilities and opacity to weight information gain
towards poorly supported Gaussian primitives of a requested object
@ObjectCentricNBV-jeong2026. That criterion prioritizes views *within* one
target. Choosing which targets to train on is a separate problem; the method
does not prescribe a product of confidence, IoU and accumulated image area for
that purpose.

// evidence:
// - @ObjectCentricNBV-jeong2026 -> docs/literature/tex-src/arXiv-Instance-NBV/ver3_rpm/3_method_ver3_rpm.tex:185-205,249-268 (confidence-weighted information gain and requested-object restriction).

=== Training allocation without changing the evaluation question

A library diagnostic allocator can group admitted observed targets by supplied
causal observation support and balance exposure across these groups, or
*strata*. It validates the support against the current decision and source
population; it does not produce the support measurements or establish their
availability for a corpus. A proposed model comparison would contrast this
allocation with uniform sampling while preserving scene diversity. Both models
would be evaluated on the same population, so the comparison would test training
allocation rather than a change in which objects count as successes.

Adaptive curricula and completed-record priority remain unselected extensions.
Their design questions are retained in the development alternatives register;
neither changes the fixed evaluation population used here.