3D matching small parts: why thin point clouds fail and the fix

Coarse-then-fine 3D matching should be the easy stage of a small-part picking cell, yet on a 22 mm grommet it locks onto the conveyor or returns nothing. The cause is almost never the camera. At 450-490 mm working distance a typical depth camera resolves a 7.5 mm thick part with room to spare; what fails is the template and the depth you feed the matcher. This guide runs the failure in the order you meet it: how much background rides along in a small-part capture, how to trim a template by hand, which depth band keeps the scene clean at match time, and when deep learning or a new camera is actually warranted.

Why does 3D matching fail on small thin parts?

On a small part the matcher is scoring a haystack: a 22 mm diameter part fills about 0.5 % of the capture area at 490 mm working distance, so more than 99 % of the points in an untrimmed cloud belong to the conveyor and the scene behind it. The share is plain geometry: the part's circle covers 0.785 × 22² = 380 mm² against an 81,600 mm² frame, and 380 / 81,600 = 0.5 %. The symptoms are consistent — the coarse stage passes with a pose on empty belt, the fine stage returns a low score or a pose centimetres off, and every capture fails the same way. Rank the causes before touching parameters:

1. Template contamination — background points survived into the matching model. This explains most failures on parts this size.
2. Scene depth range too wide — the matcher sees conveyor, guides and frame edges along with the part.
3. Template spacing or origin mismatch — the template was resampled or placed differently from the scene cloud.
4. Genuine point poverty — the rim or walls returned no points at all, which is rare with structured light at these distances.

Each cause has a two-minute test. The sections below run them in order, starting with the one that rules out buying hardware.

Is the camera resolution the problem for a 7.5 mm thick part?

No: at 490 mm a 1920-pixel-wide depth map with a roughly 340 mm field of view resolves 0.18 mm laterally, so the 7.5 mm depth step from part top to part bottom spans about 42 point spacings and is comfortably inside the sensor's reach. Work the two axes separately. Lateral spacing is field of view divided by pixels: 340 mm / 1920 px = 0.18 mm. The z step is what a top-down camera reads across the part, and structured-light cameras resolve fractions of a millimetre in z at half-metre standoffs. What the matcher really reads on a thin part is the rim circle, and at 0.18 mm spacing the 22 mm rim carries about 384 points — plenty to define a circle and its centre.

The 340 mm width is the stand-in here: take the field of view your camera's working-distance table prints for 490 mm and rerun the division — halve the distance and the spacing halves with it. Where thin parts genuinely lose points is surface, not spacing: glossy black plastic and near-vertical rim walls can return nothing on some pixels. If the rim edge is patchy in the viewer before any processing, read the close-range depth dropout guide on fuzzy surfaces and thin parts, then the fix list for missing points in a 3D camera point cloud. If the rim is solid, stop shopping for cameras: the template is your problem.

Why does remove background still leave background in the template?

Because the automatic background removal segments the capture, not the part: with the part covering about half a percent of the frame and touching the plane it lies on, the segmentation keeps a skirt of conveyor points welded to the part's outline, and that skirt then becomes part of your matching template. Coarse matching is a surface-overlap contest, and a template that is mostly plane overlaps every plane in the scene, so the coarse stage snaps the part onto the belt and the fine stage polishes a wrong pose. The tell takes ten seconds: load the template alone and rotate it. If a flat plate rides underneath the part, that plate is the matcher's favourite feature, and no parameter change will outrank it. Trim by hand instead. And when template and scene disagree in scale, origin or frame, matching fails before scoring even starts — the fine-matching input-mismatch guide walks through that separate failure.

How do you build a clean matching template for a small part?

Capture the part at the production standoff, trim the conveyor away by hand until only the part remains, and set the origin on the pick point: with a 10 mm bore the grommet's ring holds roughly 9,300 points at capture spacing, and downsampling to 0.3 mm keeps a 3,400-point template. The numbers come from the geometry: the ring between a 22 mm outer diameter and a 10 mm bore has area π/4 × (22² − 10²) = 302 mm², and at 0.18 mm point spacing one point covers 0.0324 mm².

Then trim in this order:

1. Present one good part on the production fixture at the production standoff, so the template carries the same 0.18 mm spacing as the scene clouds.
2. Capture a single frame with the production exposure and projection settings; a template from merged or enhanced frames carries edges the matcher will never see again.
3. In the point-cloud editor, box-delete everything outside the part outline, rotate the view, and delete again until no plane survives under or beside the part.
4. Run an outlier removal pass to strip the floaters left at the old boundary.
5. Set the template origin at the pick point — bore centre for a grommet, z along the part axis — so returned poses drop straight into the robot program.
6. Downsample to about 0.3 mm spacing, uniformly; decimation applied carelessly shifts points, which the downsampling distortion guide explains.
7. Validate: match the finished template back to its own source capture. The pose offset should read near zero and the score should beat every later production match by a clear margin.

Which depth filters keep the background out at match time?

Set the matching input to a depth band that hugs the part: with the top face at 490.0 mm and the part 7.5 mm thick, keep 488.0-499.5 mm, and everything beyond — conveyor included — never reaches the matcher. The band comes from three numbers: the part top at 490.0 mm, the 7.5 mm thickness putting the bottom face at 497.5 mm, and a 2.0 mm margin on each side.

The 2.0 mm margin absorbs normal sit-height variation; widen it to 3-4 mm when parts land on a sagging belt. Check the band against the worst packed scene: if two parts ever share the band, separation moves to the matching stage, which is harder work. Above the band cuts, run an outlier pass at a radius of about two point spacings to strip the last floaters, and sample the cloud uniformly — the x, y and z resolution settings guide shows how to set the spacing without distorting geometry. Keep the depth calibration honest too: a cloud scaled or clipped wrong shifts the whole band off the part, the failure the depth scale mismatch guide describes. Apply these filters to the matching input only if a downstream step still needs the full scene.

Which 3D camera suits small parts at 450-490 mm working distance?

Keep the camera you have: at 450-490 mm working distance a structured-light stereo unit resolves a 22 mm plastic part with point spacing near 0.2 mm, so change hardware only for the capture style, not for resolution. Match the style to the presentation, from the 3D sensors lines:

• One part at a time in a fixture: a projected structured-light 3D camera such as the MV-DPS200P-02 captures the whole scene in a single exposure.
• Parts moving on a conveyor: a 3D laser profile sensor such as the MV-DP4180-03P builds the cloud line by line, synced to the belt encoder.
• Coarse volume, pallet or obstacle work: depth / ToF cameras, whose millimetre-class lateral spacing at these distances is too coarse to define a 5 mm rim.

On standoff: closer tightens the spacing — halve the distance and the spacing halves — but every 3D camera has a near limit where the pattern loses focus, so stay inside the camera's rated range. The working distance and positioning guide covers that geometry, and the camera finder filters 3D sensors by type and range.

Does small-part matching need deep learning or a new camera?

No: template matching needs a repeatable part and a clean template, and a moulded grommet is both; deep learning earns its keep only when parts arrive deformed, tangled or overlapping, and a new camera only when the rim drops out in the capture itself. Decide with two questions. Is the part rigid and moulded, same geometry every cycle? Then template matching holds, and the fix list above costs an afternoon, not a training set. Do parts arrive tangled or overlapping? Then instance segmentation or trained matching is the right architecture, and the trade-offs are the same as in the AI versus rule-based inspection guide — plan it as a project, not as a patch for a dirty template.

Run either on a vision controller sized to the camera count. The MV-VC5525-256G70-NN pairs 32 GB of memory with 8 built-in light channels and PoE camera ports, so capture, matching and light timing live in one box and the pose goes straight to the robot.

How do you keep the match stable in production?

Lock the score threshold from a bench run, not from one capture: record match scores across 50 good parts, set the pass level at the lowest score you saw minus a small margin, and reject the cycle when the fine match lands below it. Then hold the rest of the chain fixed:

• Freeze capture settings — same exposure and projection state as the template session; a changed exposure shifts edges and jitter follows.
• Re-check the depth band against sit height; belt tension and fixture wear move the top face, and a 1 mm drop halves the 2.0 mm margin.
• Sanity-gate every pose before handoff: reject when z strays beyond ±2 mm of nominal or rotation exceeds the gripper's tolerance.
• Log fine-match scores per cycle; a slow downward drift flags dust, oil or mould wear on the parts, not a software fault.
• Rebuild the template whenever the part revision changes, including colour or texture, because the point return changes with the surface.

Worked in this order — trim, band, validate — a 22 mm grommet at 490 mm matches as reliably as a large casting. If a hand-trimmed template still hunts for a pose, send the raw capture, the trimmed template and one failed scene through the contact page; the fault is usually visible with the three clouds side by side. The 3D vision topic collects the rest of the point-cloud guides.