Normal profile blind human validation

Labnote 005: Normal profile blind human validation🔗

Question🔗

Does the frozen Normal UI crop profile agree with independently drawn human boxes on fresh gameplay frames?

Method🔗

• Three representative frames were selected from the fresh private Normal

validation batch before annotation.

• Each privacy-redacted frame was opened in Krita without suggested geometry.
• The operator independently drew filled rectangles on separate objectives,

skill-bar, and minimap layers.

• The bridge extracted only normalized alpha bounds after save.
• It rejected missing, empty, fragmented, or ambiguous layer masks.
• Nine independent labels were compared with the already-frozen Normal profile.
• Private images and text are not reproduced here.

Result🔗

Overall recall at IoU 0.50 was 1.000000.

Interpretation🔗

Unlike seeded overlay approval, this result provides independent geometric evidence. The frozen Normal profile is suitable as a practical crop heuristic for these three HUD regions at 1920x1080.

The minimap is most stable. The objectives panel varies in height with visible content, while the skill-bar crop is deliberately broader than the independent human boxes. Both still clear the precommitted IoU threshold in all three samples.

The sample is small and supports crop selection, not semantic detection or cross-resolution generalization.

Decision🔗

• Promote the frozen Normal geometry from provisional to independently

validated crop heuristic at 1920x1080.

• Retain objectives as a conservative maximum-height crop; consider

content-adaptive trimming later.

• Do not train a detector solely to replace this profile.
• Reuse the blind Krita bridge for Small or Large only when those profiles need

promotion.

• Larger/XL remains unsupported.