Parametric teach for geometric parts: faster vision locate

Teaching a vision system a stock part usually means photographing it, drawing a region of interest and saving a full grayscale template. For a dowel, a ring, a bushing or a rectangular blank, that is wasted work: the outline is the whole story, and the outline can be described with three or four numbers. This guide explains parametric teach — entering the shape and its dimensions so the vision system builds a small geometric model instead of storing a dense template — why the resulting locate runs faster, which part shapes qualify, and where a template is still the right tool. A worked example sizes the pixels for a 40 mm ring and shows exactly where the speed comes from.

What parametric teach is

Classic template teach saves a patch of grayscale and, at run time, slides it over the image, comparing stored pixel values at every candidate position and rotation. Parametric teach skips the photograph-as-model step entirely. You tell the system the part is a circle of 40 mm diameter, a ring with 40 mm outer and 30 mm inner diameter, a square with 25 mm sides, or a rectangle 60 by 30 mm, and the vision system builds a geometric model from those numbers alone.

Machine-tending wrist cameras were the first to make this a standard teaching option, and their makers report that teaching a stock part takes about a third of the time, because typing four numbers replaces the shoot-adjust-save loop. The model is also radically smaller, and that is what reduces locate processing time. The next two sections put numbers on both effects.

Why a geometric model locates faster

A grayscale template carries every pixel in its box and must be correlated against the image at each candidate pose, translation and rotation together. A geometric model carries curves. The matcher extracts edges in the image and fits the model's few parameters to them: a circle needs only center x, center y and radius, and since a circle looks the same at every angle, the search never sweeps rotation at all — a whole dimension of the search disappears. Edges have a second advantage: they sit where gray level changes, so they hold their position when overall brightness drifts. That is the same property that makes edge-based pattern matching stable when lighting changes, and the ambient light control techniques for pick and place apply to it unchanged.

Which part shapes qualify

The test is simple: if the outline alone defines everything the robot needs, parametric teach wins. A round spacer dropped into a chuck: yes. A shaft whose flat must face the driver: the circle finds the shaft center, and a small second model on the flat recovers rotation — the same teach-by-feature logic used for long thin parts. When the fit returns a center, that center can also anchor a coordinate system that tracks moving parts. Rotation ambiguity is a symmetry question, not a vision question; the rotation tolerance logic for symmetric parts carries over unchanged.

Worked example: pixels for a 40 mm ring on a 1.6 MP smart camera

The camera in this example is a 1.6 MP smart camera with a 16 mm lens on GigE, rated at 60 fps:

The inputs: a steel ring of 40 mm outer diameter, a typical machine-tending blank; a field of view set to 90 mm across, wide enough to absorb placement scatter and approach error; and a 4:3 pixel layout, which at 1.6 MP gives about 1,460 pixels across the frame.

The rule in words: pixel size is the field of view divided by the pixels that span it; the ring's size in pixels is its real diameter divided by that pixel size; the edge evidence is the circumference of that disk of pixels.

The arithmetic:

• pixel size = 90 mm / 1,460 px = 0.0616 mm/px
• ring diameter in pixels = 40 mm / 0.0616 mm/px = 649 px, call it 650 px
• edge evidence = π × 650 px ≈ 2,040 edge points around the outline
• template comparison: with a 20% margin the matching box is 780 px square, so a grayscale template stores 780 × 780 ≈ 608,000 pixels

Result: the ring spans 650 px against a rule-of-thumb minimum of about 100 px for a stable circle fit — below that the arc is short and the fitted center wanders — so the 1.6 MP sensor is ample. The fit solves 3 unknowns from about 2,000 edge points, where the dense template correlates roughly 608,000 gray pixels per candidate pose: about 300 times the data.

Decision: at this pixel size the outline is resolved generously, and the locate speed comes from the small model, not from more resolution. If your own diameter lands under about 100 px, narrow the field of view or shorten the working distance before reaching for a higher-resolution sensor. To run the check the other way — how many pixels a given field of view needs — use the camera resolution calculator:

Where the locate runs

For one camera and one part, the locate runs on the device. A mono sensor suits the job, because the model lives on gradients rather than color. The MV-SC5060XM-08M-WBN pairs a 6 MP mono sensor with an 8 mm lens on GigE; its rolling shutter is no handicap here, because the part sits still while the frame is grabbed.

When the cell grows to several cameras, or the locate has to share a cycle with dimensional checks and code reads, a vision controller takes over. The MV-VC3301H-128G66 is an Intel Core i3 controller with 8 GB of RAM, a 128 GB SSD, 5+ LAN ports and four built-in light channels, and it includes a VM6200Pro license. The light channels earn their place in this application: an outline only matches when it is lit the same way every cycle.

A machine-tending load cycle of 10-20 s commonly leaves a few hundred milliseconds for vision. A simple geometric locate fits inside that with room to spare; what is fast enough overall, and how to measure it on your own camera, is worked through in the smart camera processing time guide.

When a template is still the right tool

Keep the full template in these cases:

• Printed, textured or naturally varied surfaces, where the outline carries no identity and the inspection lives in the pixels. Choosing between a rule-based template and a learned model for that job is covered in the rule-based versus AI inspection comparison.
• Parts bigger than one frame, where the template itself becomes the problem — see template matching a part larger than the field of view.
• Pose defined by internal features, such as a hole pattern, that the outline cannot see.
• Low-contrast edges: metal resting on metal gives a faint gradient. Backlight the joint or add directional light before blaming the model.

Checks before you re-teach a running cell

• Repeatability: run 30 cycles with the part in a fixture and log x, y and angle. A good fit repeats to well under one pixel — about 0.06 mm at the worked example's scale. If the spread approaches a pixel, fix the edge contrast before touching the model.
• Ambiguity: with a square, confirm the 90° repetition is harmless for the grip; with a rectangle, confirm the 180° case.
• Timing: record the processing time before and after the switch in the camera's cycle readout. The gain is largest when the old template swept rotation.
• Exposure: fix it manually. Auto exposure that re-adjusts between cycles moves the gray levels under both template and model.
• Mixed stock: if one nest sees several diameters, log the fitted radius each cycle — the same fit that positions the part also identifies it.

The decision, case by case

• Round or rectangular stock whose grip ignores rotation: parametric teach, every time. Type the drawing values and move on.
• Round stock whose rotation matters: parametric model for position, small feature model for angle.
• Printed, textured or naturally varied parts: keep the template.
• Several cameras or a shared inspection cycle: run the models on a vision controller, and drive the lights from its built-in channels so the outline never changes between cycles.
• Dimensions living in CAD rather than on paper: driving parametric part files from CAD to a DXF exchange file shows how to get them into a neutral format for exactly this kind of dimension-driven set-up.

To choose the hardware around the model, start the configurator at the camera step:

The smart cameras category lists what runs the locate on the device, with the inspection camera finder filtering by interface, resolution and shutter; the vision controllers category covers the controller side. Related locate-model techniques are collected on the systems topic page, and the contact page reaches the applications team when a cell needs a second opinion on the model choice.