panel transfer baseline

Labnote 009: panel transfer baseline🔗

Question🔗

Does a frozen ImageNet-pretrained MobileNet V3 Small backbone provide enough visual structure to recognize Hero and Inventory across development batches without spending another holdout?

Method🔗

• Private input: three approved, structurally masked development batches only.
• Spent holdout: not transferred to the execution environment and not read by the experiment.
• Model: pretrained MobileNet V3 Small with frozen backbone; classifier head only.
• Augmentation: bounded translation, scale, brightness, contrast, and saturation.
• Evaluation: three leave-one-batch-out folds.
• Accelerator: RTX 4070 through a non-preemptive accelerator lease.
• Internal execution record retained; identifier withheld.

Result🔗

The job completed on CUDA with exit code 0 and no stderr. The accelerator lease was released normally.

Decision🔗

Do not freeze a new holdout or promote the model. Transfer learning materially improved over the rejected 33.3% global-feature holdout result, but generalization remains insufficient. Hero recall is complete while false positives increase in the bright third environment; Inventory performance varies by placement and batch.

Collect corrective development data emphasizing diverse no-Hero scenes, Inventory-only positions, and closed-panel negatives. Re-run development-only cross-batch evaluation before spending a new holdout.