Generative 3D shape completion: can hidden geometry be trusted?

A completed mesh can look convincing while placing a hidden wall, handle or underside in the wrong position. For teams evaluating occluded-object grasping, the purchasing question is not whether a model produces a complete shape. It is whether the remaining uncertainty permits the intended action. Generative models can suggest missing geometry, but they cannot turn an unobserved surface into a measurement. Separate sensor evidence from predictions, test the geometry that affects the grasp, and acquire another view when the distinction matters.

Decide whether the task permits inferred geometry

Generative 3D shape completion is useful when partial observations leave several plausible object shapes. A model uses learned shape patterns to propose missing surfaces, sometimes producing several alternatives rather than one mesh.

Start by assigning the output to a task:

• Object search: a rough shape hypothesis may help select the next viewing position.
• Grasp proposal: inferred geometry may suggest candidate finger positions, subject to further checks.
• Collision checking: missing observations must remain unknown, not automatically become free space.
• Dimensional inspection: a generated surface is not evidence that a hidden dimension meets tolerance.

These tasks require different acceptance criteria. A plausible underside may help rank possible grasps while being unsuitable for judging a wall thickness.

Maintain separate records for measured surface samples, inferred geometry and unobserved space. Store validity information, timestamps and coordinate frames with the original observations. Do not let mesh smoothing erase those distinctions.

Scale is a separate concern. A completion can have the correct proportions but the wrong physical size. The guide to monocular reconstruction scale ambiguity explains why a visually reasonable reconstruction does not establish metric dimensions.

How a model completes a partial shape

A depth sensor produces observations constrained by visibility, material response, sampling and noise. The completion model combines that evidence with patterns learned during training. Depending on the implementation, its output might be a point cloud, occupancy grid, surface-distance field or mesh.

A conceptual probabilistic description is:

This describes the reasoning, not an equation that every implementation explicitly evaluates. The observations constrain visible regions. The prior helps choose among shapes that could explain those observations.

Suppose a camera sees the front and upper rim of a container. Several candidates might fit: a shallow tray, a deep box or a box with a projecting rear handle. The visible data may not distinguish them. A model can favor the most familiar candidate without receiving evidence about the hidden rear surface.

Three checks therefore matter:

1. Does the prediction preserve valid observed geometry within the acquisition error budget?
2. Do alternative plausible shapes change the intended action?
3. Does a new observation contradict a previously completed region?

Repeatedly generating the same shape does not prove it is correct. A model can consistently favor the same incorrect shape, especially for objects unlike its training data.

Wireless sensing is not optical depth imaging

Shape completion is an inference method, not a sensing modality. It can operate on partial information from different sensors, but their physical limitations remain different.

Optical depth systems need suitable optical paths to the relevant surfaces. Adding a generative model does not make an ordinary depth camera measure through an opaque cover.

Millimeter-wave sensing uses radio-frequency signals. Transmission, reflection and absorption depend on frequency, material, thickness and geometry. Some barriers may permit useful signals under particular conditions; that does not establish reliable visibility through every barrier. Millimeter-wave sensing and Wi-Fi are not interchangeable specifications.

Indirect propagation can also create ambiguous observations. A system designed to interpret those paths needs an appropriate signal model and validation. An ordinary optical reconstruction pipeline cannot simply treat an indirect return as a direct surface point.

The 3D sensors category covers optical acquisition options, not a promise of through-barrier sensing. Evaluate the sensing mechanism before comparing the appearance of completed meshes.

Work through a grasp-clearance example

Consider a proposed grasp near an adjacent object. The completed shape predicts a minimum clearance of 12 mm along the approach. Your validation has established a 6 mm allowance for hidden-surface error under the tested conditions. Registration and positioning allowances are 2 mm and 1 mm respectively.

For a conservative screening calculation, subtract the allowances:

If the application requires a further 4 mm process margin, this candidate fails the screening check. A nominally open approach is not enough after uncertainty is considered.

The numbers are illustrative application allowances, not specifications for a camera or model. Their validity depends on how they were established. A percentile from a limited test set is not a guaranteed maximum error.

The calculation also assumes the relevant surfaces are represented. It cannot account for a completely missing handle or an unseen neighboring part unless those possibilities enter the uncertainty model. Check the entire finger and gripper approach volume, not just the final contact points.

Possible responses include obtaining another view, choosing a grasp on a measured surface, increasing separation or rejecting the pick. Do not reduce an allowance merely to make the arithmetic pass.

Validate hidden geometry against independent references

Evaluate completion on objects and occlusions representative of production. A collection of unobstructed, familiar objects will not characterize performance on unfamiliar parts hidden behind neighboring items.

Build reference geometry independently of the completion output. Suitable references might include verified dimensional data or a registered multi-view scan with documented accuracy. Check scan coverage: an unobserved region in the reference cannot validate the corresponding generated surface.

For a static part, acquire the reference without changing its geometry. If removing an obstruction moves the object, measure the new pose rather than assuming registration remains valid. Flexible packaging and deformable parts need particular care because handling can change their shape.

Measure both geometric agreement and task performance:

Use both surface-distance directions. A prediction can lie close to part of the reference while omitting an entire feature. A single average distance may also hide a large error exactly where a finger approaches.

Separate observed and hidden regions in the report. Otherwise, a large, accurately measured front face can dominate the average and conceal poor completion behind it.

As an illustrative reporting example, suppose 200 independent trials produce 180 correct accepted picks, 5 failed accepted picks and 15 rejections. Accepted-pick success is 180 / 185, approximately 97.3%. Acceptance coverage is 185 / 200, or 92.5%. Neither value alone describes the operating tradeoff, and these sample figures do not establish a rare-failure bound.

Split evaluation by physical object, not merely by image. Keep duplicate shapes and near-identical instances from making the held-out set artificially familiar. Include changes in occlusion, pose, surface appearance and neighboring objects.

Add evidence before adding model complexity

An additional viewpoint can resolve ambiguity that a larger completion model cannot. Choose the viewpoint to expose the surface affecting the decision, rather than collecting another similar front view.

For example, a second observation of the same visible wall may improve repeatability without revealing the rear projection that threatens the grasp. Ask which candidate shapes the next view could distinguish.

The 3D camera positioning guide helps plan acquisition geometry. Combine views only after verifying calibration, coordinate conventions and object motion. The guide to transforming depth point clouds between camera frames addresses the transformations needed to compare geometry consistently.

Object segmentation also matters. A neighboring object's depth samples can be mistaken for part of the target, giving the completion model misleading evidence. Follow the principles in selecting valid object pixels for depth calculations before aggregating object observations.

If the object moves between captures, reject stale geometry or estimate and validate the motion. More viewpoints are not automatically better when their registration is wrong.

For permanently inaccessible regions, report the remaining uncertainty explicitly. Sometimes the right response is a different fixture, a singulation step or a task that does not depend on the hidden surface.

Choose acquisition hardware without assuming completion features

Specify the acquisition problem before selecting a model. Record the object size range, required surface coverage, smallest grasp-relevant feature, allowed motion and available viewing positions. Then compare candidate sensors through the camera finder.

The MV-DLS300P-04 is listed as a laser galvanometer stereo camera. It is an optical acquisition candidate to assess against your geometry and surface requirements.

The MV-DPS200P-02 is listed as a projected structured-light camera. Its suitability likewise requires application testing rather than assumptions from the product type.

These catalog descriptions do not establish built-in generative completion, uncertainty estimation or through-barrier sensing. Treat those as separate system requirements requiring documented capabilities and representative evaluation.

Area scan cameras may provide complementary appearance images for identification or segmentation. An appearance image does not independently measure a hidden surface. If you combine image and depth streams, verify their alignment and timing rather than assuming corresponding pixels describe the same point.

Request representative raw observations before judging a completed rendering. You need to know what the sensor actually measured and what later processing supplied.

Check the acceptance rules before ordering

Use this checklist to turn an attractive demonstration into an assessable system proposal:

• Define the action. State whether the output supports search, grasp ranking, collision screening or dimensional inspection.
• Retain provenance. Require separate access to valid measurements, inferred surfaces and unknown regions.
• Test meaningful alternatives. Include objects with similar visible fronts but different hidden geometry.
• Evaluate local errors. Report performance near contact patches and approach corridors, not only across the whole mesh.
• Validate rejection. Check behavior when observations are sparse, contradictory or outside the tested object range.
• Specify another-view behavior. Define when the system requests more evidence and when it abandons the attempt.
• Track configuration. Record acquisition settings, calibration, model version and acceptance thresholds with each trial.
• Keep protection independent. Do not substitute inferred geometry for required machine safety functions.

The usual recommendation to complete the shape is wrong when the task requires verification of an inaccessible dimension or dependable exclusion of an unseen obstacle. In those cases, change the measurement arrangement or constrain the task rather than accepting a more polished prediction.

Start with the camera configuration workflow once the acquisition requirements are clear. For a feasibility review, send representative observations, reference geometry and grasp-clearance requirements through applications support. The next decision should follow the missing evidence, not the most convincing rendering.