Semantic Orientation Correction for Product Photography: A Domain-Specialized Arbitration System

Product-image pipelines routinely automate cropping, background removal, resizing, and enhancement, yet correcting the semantic orientation of a photographed product remains a distinct, largely unautomated step. Recent independent, peer-reviewed research shows that even frontier vision-language models are unreliable at recognizing image rotation. We present BlueVetaUpright, a system combining several purpose-built components behind a shared visual backbone, arbitrated by a trained decision layer, to predict both the correct 90-degree turn and a precise fine-angle correction for real product photographs.

On a large, individually hand-verified evaluation pool swept across the full rotation range, the system reaches 99.02% accuracy on clean cardinal rotations, 96.67% accuracy across the full 360-degree range (fair-scored), and 84.65% on the hardest single-answer case. A comparison against three consumer vision-language chat assistants (ChatGPT, Gemini, and Claude) on real photographs shows a decisive, statistically significant advantage against all three -- against ChatGPT, on a consolidated 394-photo sample, BlueVetaUpright leads by +15.2 percentage points on quadrant accuracy (91.1% vs. 75.9%, 95% CI [+10.4pp, +20.0pp]). All three tested assistants show a specific, recurring confusion between 90-degree and 270-degree rotations, independently replicating a finding recently documented for other frontier models in external peer-reviewed work -- a weakness BlueVetaUpright does not share.

Full interactive results, methodology, and the live tool: https://blueveta.com/upright/research