MTurk strategy

[imalsogreg]

The Turk API is tricky to work with for fully External questions. 'Other' (non-external) question type seems to allow arbitrary javascript to run in an iframe. If we can list the individual images as rows in a Task Batch CSV, and request rows one at a time without reloading the iframe source, then this would be the easiest way of hosting the face-picker widget.

[imalsogreg]

Notes from chat with @raygon (Thanks!)

• May be hard to get HIT info from query params (they come instead as 'Layout parameters' that are spliced into the dom)
• Every HIT is a page reload. So if we serve one frame per HIT, there's a lot of reflex-dom js code to reload, reparse...
• We could handle this as external questions, but there is no client side Turk API for requesting new HITs. So the external question would have to batch up several frames, and turkers would be paid on the granularity of that batch size.

So maybe the easiest thing is to bundle up frames in sets of 20, and treat each 20-frame bundle as one HIT.

[imalsogreg]

tagging (https://github.com/cbmm/tagging) now supports mechanical turk visitors. tagging has an API endpoint for serving a link to a Turk external question hosted on tagging. Now that I've played with that for other tagging projects, I think it's the easiest way forward for the face labeling task.