CONDITIONAL · COUNTERFACTUAL AUGMENTATION

Counterfactual Single-Object Removal

Preserve the existing single-object-removal experiment, but activate it only for a demonstrated background/context shortcut or over-detection failure mode.

CONDITIONALPILOT FIRSTNOT YET JUSTIFIED
Shared-audit diagnostic
Failure mode absent

Stop this branch.

Failure mode present

Run a small controlled pilot.

Choose exactly one eligible annotated object
bbox → segmentation mask → localized inpainting
Verify locality, target absence, and neighbor preservation
Remove exactly one corresponding annotation on PASS
Compare original-only vs verified single-removal pilot
Reproducible benefit?
NO

Stop; keep original training data.

YES

Consider bounded targeted scaling.

Invariant

One synthetic image removes exactly one annotated target and exactly one corresponding annotation. All retained ground truth remains authoritative.

Split safety

Generate offline and reproducibly from training only. Validation, shadow holdout if used, and test remain untouched.