User-provided mask support (for better OR)

[macmason]

It would be nice if a user could provide a mask of places to (or not to) look in an image for an object. For example, planar segmentation (cf. worldmodel) can provide good candidate objects without IDs; these candidates could then be masked in (and everything else masked out).