The plastic sheet that made spotting things in pictures much quicker
In the museum workshop, a conservator lays a clear plastic sheet over a huge tapestry. The sheet has lots of empty rectangles printed on it. Tap for “worth a look”, then nudge the edges to fit the threads. Takeaway: Faster R-CNN sped up finding things by starting with ready-made boxes and quickly fixing them, instead of hand-making boxes first.
The old way felt like two people doing separate jobs. One person crawls along the tapestry and circles loads of “maybe” spots. Only then does the second person check each circle and name what it is. The first job drags, and the second person just waits.
The twist was simple: stop drawing new circles every time. Use the same sheet idea everywhere, with a repeating set of box shapes at each spot on the tapestry. Those preset boxes are called anchors, like stencils you keep reusing. One shared look can cover small and big details, just by choosing different stencils.
At each spot, the scout does two quick things for every anchor. First, a yes-or-no tap for “is this likely something real?” Second, tiny nudges so the rectangle hugs the real edge. Clear overlaps count as good examples, clear misses as bad ones, and the fuzzy middle mostly gets skipped so the learning stays clean.
After the scout picks a short list of best rectangles, the inspector takes over. The inspector decides what each region is and tightens the box again. The time saver is that scout and inspector share the same early look, like working under one lamp with one magnifier. They got it working by training in turns until both jobs lined up.
Standing back, the conservator can stop chasing hundreds of pointless circles. A small set of well-ranked rectangles is enough, because the tap is smarter and the nudges are sharper. That’s the assumption that flipped: speed didn’t need sloppier judgement. It needed the circling step to stop being a separate chore, so tools that spot things in photos can react more quickly.