YUREI RESEARCH
← Return to research library

UI-scale stratification

zenith-003negativecomplete

Does one generic default-layout geometry describe the visible Guild Wars 2 HUD across the Small, Normal, and Large UI profiles?

blinded-reviewcomputer-visiondeduplicationguild-wars-2holdoutnegative-resultprivacy

Labnote 003: UI-scale stratification

Question

Does one generic default-layout geometry describe the visible Guild Wars 2 HUD across the Small, Normal, and Large UI profiles?

Data and review

Three private, test-only operator holdouts were evaluated:

UI profileFramesLabels
Small1236
Normal1236
Large1236

Each frame contains operator-approved objectives, skill-bar, and minimap boxes. Collection was window-bound, structurally masked, deduplicated, and separately approved for privacy and annotation. Private image and text content are not recorded here. Larger/XL was deliberately excluded by operator decision.

Generic baseline result

UI profileMean IoURecall at IoU 0.50
Small0.4182980/36 (0.000000)
Normal0.53321624/36 (0.666667)
Large0.60769024/36 (0.666667)

Per-region recall:

UI profileObjectivesSkill barMinimap
Small0/120/120/12
Normal0/1212/1212/12
Large0/1212/1212/12

Interpretation

UI scale is a material layout variable. The generic geometry happens to align best with Large and moderately with Normal, but fails Small entirely. Its objectives proposal fails every reviewed profile, so aggregate success on two regions must not be treated as detector validity.

The current reviewed boxes were generated as profile-specific seeds and then approved by the operator. They can define frozen crop-profile candidates, but evaluating those candidates on the same frames would be circular. Independent post-freeze validation is required.

Decision

  • Reject a single scale-agnostic geometry profile.
  • Preserve explicit Small, Normal, and Large strata.
  • Keep Larger/XL unsupported and do not infer it from Large.
  • Freeze profile-specific crop candidates next.
  • Evaluate each frozen profile on newly collected, separately reviewed frames
  • before promotion.

  • Continue treating all geometry as crop heuristics, not semantic detection.