Building My Own Tool Scanner — Part 16: All This, So I Could Put a Ratchet in a Drawer

Titan started with a fairly modest objection to manually measuring tools.

The result, at the point covered by this series, includes camera control, calibration targets, validation models, lighting experiments, image segmentation, vector tracing, a glass reference plane, height correction and a dedicated Gridfinity application.

All of which is a perfectly normal amount of machinery to place between a ratchet and a drawer.

This isn't the end of Titan. It's the end of the development period described in these articles, and a useful point to take stock before carrying on.

The useful thing it now does

When I started this, I wasn't trying to invent a whole new category of machine. I wanted to stop spending so much time drawing the shapes of tools that were already sitting in front of me. Put a tool down, get an outline, make an insert, move on to the next one. That was the plan. Somewhere along the way, the plan acquired a calibration system, a vector pipeline, a dedicated CAD generator and a surprisingly long list of camera experiments. All entirely reasonable, obviously. 😂

The accepted setup uses the Samsung A3 with controlled capture settings, a glass working plane and the complete dense dot target. Production mapping uses the observed V2 grid to build the B-spline rectification.

The scanner captures a fresh image, preserves its identity and diagnostics, detects the visible geometry and produces nominal SVG profiles. I can choose the outside pocket silhouette or retain detected internal holes.

The generator then handles the manufacturing decisions: layout, clearance, depth, finger access, the reference-derived insert and baseplate geometry, splitting and export.

For raised objects, I can supply height information to the optional correction. Physical X/Y measurements also remain an explicit tool where appropriate.

That is much closer to the original aim than manually recreating handle curves in CAD.

It isn't a claim that the camera knows everything about the object. The workflow has become useful partly because the things it can't know are more clearly identified.

There wasn't one accuracy problem

The useful difference now is that I can see the separate jobs much more clearly. The camera captures the image. Calibration says where points lie on the glass plane. The detector decides which pixels belong to the tool. The tracer turns that boundary into geometry. The generator decides how the insert should be made. It sounds obvious written out like that, but quite a few of the development headaches came from assuming one of those stages could quietly compensate for another.

Several quite different failures initially looked like the same complaint: the size was wrong.

Sometimes the detector selected an interior boundary. Sometimes a reflective edge disappeared. Sometimes a validation endpoint was snapped to the answer. Sometimes the object was elevated above the calibrated plane.

And, eventually, the photograph was of an entirely different tool.

Changing the calibration for all of those would have produced a remarkably busy calibration file and a very confused scanner.

Separating the stages made the investigations more useful. Capture establishes the image. Calibration maps the reference plane. Segmentation chooses the visible boundary. Vectorisation represents it. CAD and the printer turn it into a manufactured part.

Each stage can be wrong in its own way.

The failed branches earned their place

Looking back, some of the most useful results were the failures. A validator that snapped endpoints to expected positions taught me to distrust an answer that looked too perfect. Lens changes taught me that sharpness alone doesn't make a scanner better. The phone experiments taught me that more pixels aren't a substitute for a stable physical setup. And the cached-image fault taught me that all the calibration in the world is useless if I'm processing the wrong photograph. I'd rather have those lessons written down than repeat them six months from now.

The global grid-phase correction rejected genuine evidence. The replacement lenses didn't improve the complete system. The Z3 route was inconvenient. Smoothing could remove tips, while retaining every contour point made CAD unpleasant.

The A3 didn't become the final camera in one clean step either. There was a temporary return to the C920, mixed taped-sheet results and a mechanical change to the scan plane before the accepted glass configuration.

Those experiments aren't embarrassing footnotes to remove from the account. They explain why the current choices exist.

They also keep a future change from looking obvious simply because I've forgotten the failed attempt that preceded it.

The numbers still need their labels

The Gridfinity side deserves its share of the story too. A brilliant SVG isn't the finished product. It still needs to become a pocket with the right clearance, depth, access and printable geometry. That was why I ended up building the generator alongside the scanner. I hadn't set out to write two programs, but doing only half the job would still leave me sitting in CAD repeating the same operations I wanted to avoid in the first place.

The series contains some good calibration and object results. It also contains maxima that are less flattering than the averages, a changed detector acceptance rule and a particularly pleasing ratchet length that mustn't become an advertised global specification.

The useful position now is fairly specific.

None of that is improved by averaging unrelated experiments into one attractive accuracy number.

The same applies to printing. A successful ratchet insert is valuable practical evidence. A later print described as “in progress” doesn't acquire a successful outcome because the earlier one fitted.

Good enough for a job isn't finished forever

I also don't want to finish this series by pretending every tool now measures perfectly, everywhere, under every condition. It doesn't. There are real limits to a single overhead camera, especially with reflective edges and tools that change height across their footprint. The point is that those limits are much better understood now, and the system can do useful work inside them. That's a stronger position than claiming an accuracy figure the evidence doesn't support.

The original target was useful tool storage, not a laboratory measurement instrument.

That remains a helpful way to decide what deserves attention. A change is interesting when it improves the reliability, convenience or manufactured result, rather than merely supplying another decimal place to admire.

There is still development to do. Irregular heights, difficult reflective edges, small internal features and awkward CAD fillets remain meaningful areas to investigate. So does making the ordinary capture and operating workflow dependable.

The better historical record matters here as well. Future changes can be compared with actual configurations and retained evidence, rather than a vague recollection that one camera or one calibration was somehow better.

The next articles belong to the next work

So this is a stopping point for the articles, not for the project. There are still tools to scan, inserts to make, awkward edge cases to investigate and hardware changes I'd like to try. The difference is that I've now got a proper record of what worked, what didn't, and why I made the choices I did. All of which should make the next round of tinkering slightly less chaotic. No promises, mind.

I haven't run out of things to improve, and this retrospective isn't a declaration that Titan is complete.

It's a record of the point reached so far: a working, useful scanner and manufacturing workflow, together with a much better understanding of its limitations.

Future articles can follow the next experiments as they happen. If a new approach replaces something described here, that will be another stage in the development rather than a reason to pretend the old stage never existed.

The ratchet can have its drawer.

Titan can keep developing.