The stagehand who could name a prop without looking
In the theatre store room, my torch died. I grabbed a prop with tiny raised bumps and had to name it fast, just by touch. My fingers bounced around at random, then I held on to the few bumps that felt most telling. Takeaway: a handful of strong touches can still reveal the whole shape.
People used to treat scattered 3D dots like a nuisance. So they’d force them into neat boxes, or turn them into a few flat pictures first. That’s like making me draw the prop on graph paper before I’m allowed to guess. You lose fine bits, and you do extra work.
PointNet goes straight from the dots. It checks each dot with the same little “finger test”, turning it into a few simple signals like edge-ish, corner-ish, flat-ish. Then it uses an order-blind gather: for each signal, it keeps only the strongest one seen anywhere, so shuffling the dots changes nothing.
One snag: the same prop can be turned in your hands. PointNet adds a step that tries to straighten the dots into a familiar pose first, like me turning the prop until it sits right. It’s also kept on a tight lead, so the straightening acts like a clean turn, not a squash or stretch.
That “keep the strongest” trick has a funny side effect. Only a small set of dots can ever win as the strongest signals. They’re like the few fingertip touches that actually convinced me. Drop lots of other dots, or add noisy ones that never beat the winners, and the overall guess often stays steady.
Once it has a whole-object summary, it can do two jobs. It can name the object, like “chair” or “mug”. Or it can point to parts, by pairing the whole summary with each dot’s local clue, like “this bit is the handle” and “this bit is the leg”, even in a full room scan.
Back in the store room, I didn’t need a perfect drawing or a perfect touch order. I just needed a few standout bumps. That’s the flipped assumption: messy 3D dots don’t have to be forced into tidy grids first. Treat them like an unordered handful, keep the strongest clues, and you can still make solid sense of the shape.