Do not use text-ROI area alone as a capture-quality gate

General lesson: for scene-text object detectors, the ground-truth object and the product’s capture-acceptance policy must remain separate. Annotate a stable semantic unit (for example, the complete required label plus its value and directly associated optional field), excluding logos and unrelated context. A true but tiny object remains a positive detector label even when the product may ask for a closer retake. Full-image ROI area is only a distance proxy: thin single-line text, multiline blocks, source aspect ratio, and boxes inflated by logos/background can produce the same area with very different readability. Apply any context padding after detection, and calibrate product rejection on a source-disjoint, human-adjudicated set using downstream exact-field retention plus text/glyph scale—not area alone.

Applies when: training or gating object detectors for small text regions embedded in larger real-world images. Keywords: ROI boundary text detector area threshold readability padding annotation leakage