# Manual review comparison

Snapshot: 2026 larger screenshots, revision 165. Raw reports are unchanged. Current saved labels, rather than every historical drag, are compared. The snapshot and detailed comparison.json remain local.

Nine photos have 121 decisions: 101 corrected positions and 20 obscured joints. The unpaired 2025 originals have no saved labels yet, so these reviews cannot establish whether higher-resolution photos improve accuracy.

Of the 101 corrected positions, 79 have a usable original 3D projection and 22 have a usable original 2D point. These denominators mix detector-specific joint vocabularies and missing/low-confidence detections; they are not directly comparable recall scores.

The matched subset contains 19 manually reviewed joint locations for which both detectors have usable original points. Median source-pixel discrepancy from manual placement: 3D 30.05 px; 2D 13.02 px. All sources here are similarly sized screenshots. This small, selected subset favors 2D localization but does not establish overall superiority.

Largest 3D discrepancies:
- P05-reference right ankle: 209 px; right knee: 192 px.
- P06-subject right ankle: 168 px; right knee: 156 px.
- P06-reference right knee: 153 px.

These warrant checking anatomical side assignment and occlusion rather than applying a global coordinate offset. Manual placements are reference labels, not independently verified anatomical ground truth. Obscured and uncertain labels are excluded from positional error. A corrected point does not recover depth.

The labels can support controlled comparisons of crop strategies, detection modes, or future models using held-out photos. The current application does not train or update Apple's detector from edits. More reviewed high-resolution photos and visibility decisions would strengthen evaluation; an expert-rider match is unnecessary for testing joint localization.

Reproduce with:
`python3 scripts/compare-reviews.py artifacts/review-comparison/revision-165.json`