Mind Map

Content

Instructions for Creating Target Object Point Cloud Templates

I. Prerequisites

Prerequisite	Definition	Notes
Target Object pose variation	Limited poses (countable): only fixed faces of the Target Object can be seen (for example, the front side), and the valid Point Cloud area is the region of that face that remains stably visible. Arbitrary poses: the Target Object may appear at any angle, and the valid Point Cloud area is the region that can be imaged stably in more than 80% of poses (excluding deep holes, overexposed areas, grooves, and other parts where Point Cloud is easily missing).	Ordered scenes: The template should depend only on the Point Cloud of the fixed visible face to avoid matching failure caused by invisible faces. Random scenes: Stable visible regions must be determined through multi-view scanning or CAD simulation. Exclude regions where data is easily missing (such as reflective surfaces and deep concave structures) to avoid relying on unreliable Point Cloud. Confirmation method: Collect Point Cloud data under different poses, count the visibility frequency of each region, and screen out the valid Point Cloud area.
Camera Point Cloud imaging quality	The quality and completeness of the Point Cloud data obtained after the Target Object is recognized by a 3D Camera (such as structured light or binocular vision).	Missing Point Cloud Common causes: Target Object properties (high reflectivity or light-transmitting materials prevent the laser from returning). Ambient light interference (strong lighting causes sensor saturation). View occlusion (self-occlusion of the Target Object or occlusion by external objects). Optimization strategies: Adjust the light source: use diffuse light or polarized filtering to reduce reflections. Local template: use Point Cloud data from visible areas with better quality as the template. Discontinuous Point Cloud depth and severe noise Common causes: Target Object properties (multi-layer structures, transparent or translucent materials). The Camera is outside the disparity range (measuring too near or too far causes Point Cloud distortion). Stacked-object occlusion (for example, stacked workpieces cause Point Cloud mixing). Optimization strategies: Filtering and denoising: use filtering and outlier removal functions to optimize Point Cloud quality. Depth calibration: ensure the Camera collects data at the optimal working distance. Layered analysis: extract Point Cloud by region for stacked workpieces to avoid mismatching.

II. Target Object Characteristics

Attribute （higher priority means more important）	Definition	Notes
Size	The physical size of the Target Object (volume, area) affects Point Cloud collection density and template computation complexity.	Large Target Objects: It is recommended that the number of points in the Point Cloud template not exceed 50,000 (whether to downsample or cut at the original density and use it independently as a fine-registration template should be decided according to accuracy requirements) to avoid wasting computing resources. Prefer regions with obvious features (such as edges and holes) for dense Point Cloud sampling. Purely planar regions can use lower Point Cloud density to improve registration efficiency. Small Target Objects: Ensure good Camera imaging quality so that Point Cloud details are complete. If there are too few visible Point Cloud points (fewer than 200), lower the Camera appropriately to improve Point Cloud density.
Self-occlusion	The structure of the Target Object itself causes some regions to fail to form Point Cloud under certain viewpoints (for example, multi-layer nesting, stacked workpieces, complex curved surfaces, and so on).	Occlusion edges are prone to Point Cloud breaks or noise. It is recommended to directly use the Point Cloud processed by the PickWiz system as the Point Cloud template. Stable features that are insensitive to occlusion (such as outer contours and fixed hole positions) should be selected as the main body of the template, while time-varying occluded areas should be ignored. If serious occlusion exists under a single viewpoint, local Point Cloud must be collected from multiple viewpoints to create templates.
Three-dimensionality	The degree of depth variation of the Target Object in 3D space (for example, height differences, stepped structures, curved undulations, and so on).	The regions that can form Point Cloud should be used as the criterion for selecting valid template regions, avoiding interference from Point Cloud on the back side. Edge and corner regions can use higher Point Cloud density; curved regions can retain the Camera Point Cloud density; planar regions can be sparsely sampled to balance accuracy and efficiency. Use 50,000 points as the upper benchmark and adjust template density on that basis.
Symmetry	The Target Object has a symmetric structure (for example, bilateral symmetry or rotational symmetry).	Keypoint calibration: if a keypoint model is used, the Point Cloud template and CAD center point must be aligned strictly to avoid registration offsets caused by mirrored errors.
Concave shape	The Target Object surface has recessed structures (such as inner holes or grooves), which may lead to missing Point Cloud or unstable imaging.	Prioritize the outer contour: extract the complete visible outer contour first as the main body of the template. Consider stability: Side Point Cloud: assess stability under different viewpoints and choose the most stable visible region. Bottom Point Cloud: visibility is usually poor, so dependence on it should be reduced or limited to key regions. Simplify the template: if the Point Cloud quality in groove areas is poor, create the template using only the edge contour plus a small number of feature points.
Convex hull shape	The Target Object surface has recessed structures (such as inner holes or grooves), which may lead to missing Point Cloud or unstable imaging.	Reflection issue: convex curved surfaces easily cause edge reflections, leading to unstable Point Cloud imaging quality. The template should prioritize complete Point Cloud regions. Prioritize strong features: if the overall Point Cloud quality is poor, keep only highly distinctive features (such as marked points and edges) for registration.
Planar type	The Target Object surface is approximately planar and lacks obvious internal features.	Strengthen edges: concentrate Point Cloud distribution on edges to enhance registration guidance. Compensate for missing inclined Point Cloud: If the Target Object tilts and causes blurred Point Cloud imaging at the edges, use a keypoint model to assist registration. Or optimize Point Cloud quality through multi-angle lighting.
Dynamic update status	/	For Target Objects that are prone to wear or deformation, re-evaluate changes in occlusion and three-dimensionality regularly and update the template.

III. Basic Principles

When using model-based matching in industrial Robot 3D vision, feature selection should follow the principles below
Characteristic	Description
Salience	Select features that are prominent and easy to identify on the Target Object surface, such as edges, corner points, textures, and concave or convex surfaces. These features should have strong discriminability to enable fast detection and matching in the scene.
Stability	Features should remain relatively stable under changes in lighting, viewpoint, and so on. Even if the Target Object rotates, scales, or deforms, these features should stay relatively unchanged.
Uniqueness	Features should be distinctive enough that they are not easily confused with features from other Target Objects or the environment. This helps improve matching accuracy and reliability.
Density	Distribute feature points as densely as possible on the Target Object surface to increase matching redundancy. Even if some feature points are occluded or lost, matching can still be completed using other feature points.
Repeatability	Select features that can be detected stably under different viewpoints and lighting conditions. This ensures that the features can be recognized accurately in various scenes.
Computational Efficiency	The extraction and description process for features should be as efficient as possible to meet the real-time response requirements of industrial Robots. Prefer features with relatively low computational cost when possible.

IV. Notes for Specific Scenes

The general considerations for Point Cloud template creation under different industrial scenes and technical approaches are as follows.

Scene	Scene Image	Point Cloud Image Scene screenshot Point Cloud file	Template Direct extraction image (file + image) Final template (file + image)	Notes for Point Cloud Templates: prerequisites × Target Object characteristics - basic principles - object selection - specific form	Recommended responses to the notes above
General Ordered				Prerequisites: Target Object characteristics: large long-strip piece (possibly deformable), planar Target Object. Basic principles:	Prerequisites: Target Object characteristics:
General Random			Direct extraction image	Prerequisites: Target Object characteristics: self-occlusion, concave shape (with round holes). Basic principles:	Prerequisites: Target Object characteristics:
Surface-Type Ordered				Prerequisites: Target Object characteristics: three-dimensionality (height differences and curved undulations). Basic principles:	Prerequisites: Target Object characteristics:
Surface-Type Ordered	Process error causes the spatial position of the small can to fluctuate up and down.			Prerequisites: Target Object characteristics: three-dimensionality (height differences and curved undulations). Basic principles:	Prerequisites: Target Object characteristics:
Surface-Type Random				Prerequisites: Target Object characteristics: symmetry (rotational symmetry), self-occlusion. Basic principles:	Prerequisites: Target Object characteristics:
General Random				Prerequisites: according to scene placement and Target Object characteristics, front and back templates should be used. Create separate front-side and back-side Point Cloud templates for the Target Object, and use overall keypoints for the Target Object. Target Object characteristics: convex-hull type; convex curved surfaces easily cause reflections and unstable Point Cloud imaging quality. Basic principles:	Prerequisites: use PickWiz to generate front-side and back-side Point Cloud templates separately (current viewpoint), and generate overall keypoints for the Target Object. Target Object characteristics: use dual scene Point Cloud templates, and the templates should fit the front and back sides of the CAD model.
Surface-Type Positioning Assembly (matching only)				Prerequisites: Target Object characteristics: Basic principles:	Prerequisites: Target Object characteristics:
General Positioning Assembly				Prerequisites: Target Object characteristics: Basic principles:	Prerequisites: Target Object characteristics:

Previous pagePoint Cloud Template Creation Guide

Next pageScene Object Usage Guide