Hardware shops still count nuts and bolts by hand, one at a time. Counting Station lets a shopkeeper tip a handful onto a sheet of paper and get the count from a phone camera, and it speaks up when it isn't sure.
No scale, no special machine, no training data. The mat is an ordinary A4 sheet with four printed markers in its corners. The camera can be held at any angle: the markers flatten the photo into a top-down view and turn pixels into millimetres.
That one trick does most of the work. Once everything is measured in real units, the size of one screw is a physical fact rather than a guess, and a blob twice that size is two screws.
Find the markers (three are enough if a hand covers one) and warp the photo onto the mat.
Black parts and shiny steel need opposite treatment, so it measures first and picks the right method.
Dark parts: keep only their near-black core, so shadows and grease drop out. Bright parts: keep anything with crisp edges, so glare doesn't break them up.
Once by area against the size of one part, once by splitting clumps along their narrow necks. The two answers are reconciled.
Every count comes with a range. When a clump is unclear it says so, shows where, and asks you to spread the parts.
Lay a few loose pieces down once and it learns that part's exact size, which removes most remaining errors.
For each photo the station can lay out every intermediate image side by side, so when a count is wrong you can see exactly which step went wrong instead of guessing. Here, 8 shiny steel screws on the mat.
A desk lamp throws hard shadows, and a simple threshold glues each part to its shadow and its neighbours. The final version divides every pixel by how bright the paper around it should be, so paper, shadow and metal separate cleanly, and only the near-black core of each part is kept.
Alongside the classical pipeline sits a small learned detector, YOLO11n-OBB. Nobody drew a single box for it: it learned from thousands of synthetic scenes built by cutting real screws out of a handful of photos and pasting them onto empty mats, with fake shadows, glare and blur.
Each is strong exactly where the other is weak: counting by area handles tight clusters of small screws, the detector handles long screws lying across each other. So when they agree, the count can be trusted, and when they don't, the station asks for the parts to be spread out before it commits.
Of counts were correct whenever the two methods agreed, so agreement is a trust signal you can act on
The most it was ever off on a handful of screws, once it had been shown the part a single time
Per photo on one laptop CPU core, no GPU needed
Every number regenerates from a frozen, checksummed snapshot of the code and data.