BlueVeta
An AI team focused on visual processing of product images. It developed the dedicated rotation detection model BlueVetaUpright, aiming to solve vertical visual tasks such as automatically straightening product photos.
Timeline (last 90 days)
BlueVetaUpright achieved 91.1% accuracy on a four-quadrant rotation detection task with 400 real product photos, while ChatGPT achieved 75.9%, a gap of 15.2 percentage points (95% confidence interval [+10.4, +20.0], Bootstrap resampling 10000 times clustered by source photo).
The experiment used 400 real product photos, split into two independent batches of 200 each, with no shared source photos between the two batches, to rule out the risk of data leakage.
The BlueVeta team explained that the fluctuation in ChatGPT's accuracy between the two batches is because the consumer chat interface does not provide a fixed model snapshot guarantee, and backend routing may change between sessions.
The confidence interval of the test results for the first group of 200 photos touched zero and did not reach statistical significance; the team therefore built a second independent batch to verify.
The complete experimental method, bucket breakdown and limitations are published at blueveta.com/upright/research, and the research paper is archived on Zenodo (DOI: 10.5281/zenodo.22975679).
All Facts (5)
BlueVetaUpright achieved 91.1% accuracy on a four-quadrant rotation detection task with 400 real product photos, while ChatGPT achieved 75.9%, a gap of 15.2 percentage points (95% confidence interval [+10.4, +20.0], Bootstrap resampling 10000 times clustered by source photo).
50%UnverifiedThe experiment used 400 real product photos, split into two independent batches of 200 each, with no shared source photos between the two batches, to rule out the risk of data leakage.
50%UnverifiedThe BlueVeta team explained that the fluctuation in ChatGPT's accuracy between the two batches is because the consumer chat interface does not provide a fixed model snapshot guarantee, and backend routing may change between sessions.
50%UnverifiedThe confidence interval of the test results for the first group of 200 photos touched zero and did not reach statistical significance; the team therefore built a second independent batch to verify.
50%UnverifiedThe complete experimental method, bucket breakdown and limitations are published at blueveta.com/upright/research, and the research paper is archived on Zenodo (DOI: 10.5281/zenodo.22975679).
50%