Machine vision-generated labels for Commons images will be requested from one or more MV providers. We need to answer the following lifecycle questions about this data:
| Status | Subtype | Assigned | Task | ||
|---|---|---|---|---|---|
| Resolved | • Ramsey-WMF | T225964 [SDC] Build a depicts tag suggestion tool that is powered by machine vision platforms | |||
| Resolved | • Mholloway | T226119 Build middleware to utilize machine vision API for structured data on commons depicts tag suggestion tool | |||
| Resolved | • Mholloway | T227353 Define the lifecycle for MV-generated depicts suggestion data |
From Adam over email:
are there use cases for holding onto the labels for some interval post the first human verification? My gut intuition is we probably want to stash the latest retrieved label data indefinitely even if the original data for the paid services is only internet exposed until a human verification (plus some lag maybe to deal with risk of vandalism?), but in addition to just needing to deal with up front storage planning.
I'll just add to this one possible future use case we've discussed before: when/if WMF eventually has our own homegrown system for this, it could be useful to have the old labels from V1 to compare/contrast against and judge the quality of our new model(s).
I'm in support of permanent stashing.
It's a subtler matter, but the question of refresh frequency is orthogonal.