Part size vs locate time: feature count, not size

You teach a 40 mm ring on the bench and the locate returns in well under a second. Swap in a 120 mm plate, teach the whole surface, and the same call now takes seconds. The reflex is to look for a footprint threshold, some size past which the camera slows down. There is no such threshold: a pattern matcher bills the features it must test, not the millimetres a part spans. Size enters only because bigger parts usually carry more features and fill more pixels. What follows is the mechanism, the arithmetic that turns part size into locate time, the anchors to plan with, and the teach choices that buy the seconds back.

What a matcher does with your seconds

A locate is a search. The matcher extracts candidate features from the live image — corners, edges, blobs — and scores them against the features stored in the taught model at every plausible position and rotation. Three quantities set the work per shot:

• Model features: what you taught, fixed at teach time.
• Scene candidates: features the image offers, including anything the background contributes.
• Search area: the position window and rotation range the matcher sweeps.

The relation is a planning proportionality, not a data-sheet value. It explains both effects you meet on the floor: a cluttered background raises k by feeding the matcher false candidates to reject, and a bigger taught region raises N_model directly.

The two anchors to plan with

For an embedded smart camera locating on the device, plan with roughly 0.4 s for a simple model — a ring on a plain background — and 1 to 5 s for a part with hundreds of features. The processor sets the constant and the feature count sets the scaling, so time one cycle on your own hardware before locking a cycle budget.

Every speed-up acts on one of the three factors: fewer model features, fewer scene candidates, smaller search area.

How part size turns into locate time

Feature density is the bridge. On a machined or textured surface, detectable features spread roughly uniformly, so the count inside a taught region grows with its area. At the same mm/px, doubling a part's linear size quadruples its area, its feature count and its pixel count together — and teaching the whole surface picks up all of it. Pixels follow the same law: area in mm² divided by (mm/px)² gives pixels on part.

Rotation range belongs in the same budget, because the sweep is part of A_search: a ±180° sweep costs more than a ±10° one.

Worked example: 0.4 s becomes 4.6 s, then a few hundred ms

One camera, one field of view: 300 × 200 mm on a 6 MP frame, taken as 3000 × 2000 px, so 300 mm / 3000 px = 0.1 mm/px.

1. Baseline. Ring, 40 mm diameter, plain background, taught whole. Area = π/4 × 40² = 1257 mm². Locate: ~0.4 s.
2. Bigger part, same teach habit. Plate, 120 × 120 mm, similar surface. Area = 120 × 120 = 14,400 mm². Ratio 14,400 / 1257 = 11.5, so expect 0.4 s × 11.5 ≈ 4.6 s — the top of the complex-part band. The plate spans 1200 × 1200 px, 1.44 million of the frame's 6 million pixels.
3. The fix. Teach a 30 × 30 mm region around a distinctive notch: 900 mm², or 300 × 300 px. Model features drop 16-fold (14,400 / 900). Clamp the search to a 60 × 60 mm window around the fixture position: 3,600 mm² instead of the full 60,000 mm² frame, 16.7× fewer sweep positions. The proportionality predicts 4.6 s / (16 × 16.7) ≈ 17 ms.
4. The floor. Real software carries fixed cost no model change removes — image transfer, feature extraction over the window, pose output — so the practical result is a few hundred milliseconds, near the simple-model anchor. Seconds traded for fractions of one.

Sizing the sensor so the notch keeps enough pixels is the same conversion run in reverse; the calculator below does it and opens the camera finder filtered to sensors with enough pixels.

Where hundreds of features come from

Count features, not holes. Every hole contributes its edges, every engraved character its strokes, every knurl a dense band of corners. A plate with a 10 × 10 hole pattern carries at least a hundred strong edge features from the holes alone; add machined texture and you are firmly in the hundreds band. Cast, forged, wooden and fabric surfaces raise the density everywhere and reward the smallest teach region that still works.

If parts of a family differ by hole count, that difference is itself feature-based: keep one small ordered model per type, as in differentiating parts by hole count with ordered locate models.

The background: free speed or a feature factory

A plain, matte, uniform background adds no candidates. Any geometric shape inside the search window — a tray edge, a printed carton, a bin wall, the previous part left in view — is a candidate the matcher must test and reject, and each one costs time. The magnitude is the same order as the part: one tray edge is a handful of candidates, a printed or textured backdrop face can add hundreds. If the background moves between shots, it can also hand back a confident pose for something that is not the part.

Teach against the same backdrop you run on, keep it unpatterned and matte, keep fixtures out of the taught region, and give the part the same presentation every cycle. A camera stand with a glass tray or a matte plate holds that presentation steady.

Teach less, keep the pose

A locate returns x, y and rotation, and a small model can deliver all three. Position comes from any distinctive group; rotation precision depends on the baseline — how far the features sit from the model's reference point:

At 0.1 mm/px, features spanning 200 px with a 0.3 px centroid jitter give 0.3 / 200 = 0.0015 rad ≈ 0.09°. A 30 mm teach region spans 300 px at that scale, so sub-0.1° rotation from a small region is normal, provided the features inside are distinctive and well separated.

Three teach rules that hold across pattern-matching software:

• Two or three well-separated feature groups — about a dozen strong features — fix the pose. Hundreds mean you taught texture the pose never uses.
• On a symmetric part, teach the one feature that breaks symmetry, a chamfer or an off-centre hole, or rotation snaps to mirror angles.
• Clamp the search window to the expected position plus the fixture's placement tolerance, and set the rotation range to what the process actually produces.

The same principle applied from other directions: teach long thin parts by feature, not the whole part, and parametric teach for geometric parts.

Before you blame the part size

The matcher is one phase of a snapshot cycle. Trigger, exposure and image transfer sit before it, and none of them shrinks when you shrink the model. If a plain test card also locates slowly, the time is in focus, exposure or capture timing: start with fixing focus and exposure to cut snapshot time and trigger call versus exposure time. If the robot reaches the snapshot pose late, you are timing the wait, not the matcher — see why the robot must return to the snapshot position first. For the full budget, where 5 to 7 seconds of a wrist-camera locate go breaks a real cycle down.

Checks before you release the faster model

1. Record the model's feature count before and after; your vision software reports it at teach.
2. Time 30 consecutive locates and keep the worst, not the mean.
3. Repeat at the extremes of the part's rotation range — that is where a clamped window is tested hardest.
4. Check pose spread over the same 30 runs: x and y in mm, rotation in degrees, against the process tolerance.

The decision, case by case

When the part does not fit the frame at all, teach-size advice does not apply: read locating a large part when the field of view is too small and template matching a part larger than the field of view. If you are still deciding whether the part can be located at all, start from the vision locate feasibility checklist.

Matching the device to the presentation

Vision sensors and smart cameras run the locate on the device; a camera-plus-library set-up runs it on a PC. Match the class to the case:

• Single part on a plain tray: the MV-SCB007EC-12S-WBN vision sensor pairs a 1.6 MP sensor with a 12 mm manual-focus lens and runs the whole locate on the device.
• Wider field of view where features must keep their pixels: the MV-SC5060XM-08M-WBN smart camera pairs a 6 MP mono sensor with an 8 mm M12-mount lens over GigE.
• Fixed distance, fixed presentation: the MV-SCA002M-03S-WBN-SR vision sensor focuses its 3 mm lens at 240 mm, and its 0.3 MP sensor keeps the frame small.

The stand matters as much as the sensor here: it fixes the working distance, and a slide-table type lets you re-frame without re-teaching. The configurator below starts at the camera step, and the camera finder filters the range by resolution, interface and shutter if you already hold the numbers.

Measure one locate on your part, teach the smaller region, clamp the window, and compare. If the seconds survive both changes, send the model's feature count and the phase timings through the /contact/ page — that pair separates a matcher problem from a capture problem fast.