Why overlays are harder than text replacement
Recognizing text is only one stage. A useful overlay must track where that text appeared, how long it remains visible and whether translated content fits the same visual region.
Layout is a dynamic constraint
Translation often changes sentence length. A fixed-height box can clip long translations; a box sized only once can leave excessive empty space when the next sentence is short. Measuring rendered text and respecting safe screen bounds helps.
Timing matters as much as recognition
Fast dialogue can disappear between OCR samples. More frequent scans may help, but they also increase device work. A useful design balances change detection, sampling cadence and deduplication rather than maximizing scans blindly.
What we would validate
Test with mixed font sizes, animated dialogue, different screen resolutions, multi-line sentences and rapid transitions. Measure missed text, overlay legibility and end-to-end response time separately.
Thoughtful engineering starts with accurate problem framing, clear constraints and repeatable experiments.