PaperlayerView exact revision on arXiv

Source-linked AI summary

Prior Availability in Industrial Visual Sim-to-Real: A Review of CAD-Guided and CAD-Unavailable Regimes

Chenxi Tao, Seung-Kyum Choi

arXiv:2605.30581v2cs.CVcs.AIcs.RO

TL;DR

Industrial visual sim-to-real spans different available priors and deployment evidence, raising the question of which prior can ground deployment decisions. This review organizes methods by prior availability and finds that render count alone does not close transfer, while test-time CAD adds geometric verification.

  • Problem

    Industrial visual sim-to-real lacks an organizing account of how CAD availability changes the priors, mismatches, and test-time verification available to deployment systems.

    Source: §1 · p. 2

  • Method

    The review organizes industrial sim-to-real methods by CAD availability, treats boundary priors as intermediate cases, and uses empirical anchors across detection, pose, and anomaly inspection.

    Source: §1 · p. 4

  • Results

    The anchors show render count alone does not close transfer (mAP50:95 decreases 0.1521 to 0.1287 as PBR renders rise 5k to 50k); source design, capacity, calibration, and test-time CAD also matter.

    Source: §6.2 · p. 25, §6.3 · p. 28

  • Takeaways & Limitations

    The prior-strength rubric offers a reproducible way to state what evidence a method carries across the domain gap without treating it as a universal benchmark.

    Source: §3 · p. 6

    CAD-based verification depends on usable proposals, object identity, calibration, depth or mask information, and refinable pose hypotheses, and can weaken under occlusion, symmetry, reflective materials, or clutter.

    Source: §4.3 · p. 12

Abstract

from arXiv · show

Industrial visual sim-to-real is often described as transferring from synthetic images to real images, but industrial deployment usually involves a broader mismatch between available evidence and required decisions. A system may be built from CAD renderings, simulated RGB-D observations, normal reference images, synthetic defects, pretrained feature spaces, or language prompts, yet deployed under different sensors, lighting, materials, fixtures, calibration, production variation, and rare defect modes. This review reframes industrial visual sim-to-real as a domain-gap problem organized by prior availability. We distinguish CAD-available settings, where explicit object geometry can support rendering, calibration, pose estimation, segmentation, and test-time geometric verification; CAD-unavailable settings, where geometry is replaced by normal-reference appearance, feature distributions, teacher-student residuals, synthetic anomaly assumptions, foundation features, or vision-language priors; and boundary-prior settings, where approximate models, templates, reference views, or semantic correspondences preserve only part of the CAD role. This framing connects CAD-based detection and 6D pose-estimation literature with industrial anomaly and surface-inspection literature that is usually reviewed separately. To make the taxonomy concrete, we use empirical anchors on T-LESS/BOP, MVTec AD, and VisA. The anchors show that CAD render count alone does not close transfer; source-distribution design, detector capacity, and small real calibration can matter more. They also show that CAD at test time creates a distinct verification channel through mask, pose, and depth consistency, whereas CAD-unavailable inspection relies on calibrated normality and feature deviation. The review therefore argues against a single cross-task leaderboard and instead asks what prior grounds the deployment decision.

1 Introduction

Industrial visual sim-to-real is a deployment domain-gap problem: systems must make decisions in factory conditions that differ from the evidence used to design, train, or configure them. The review organizes this gap around available priors, distinguishing geometry-based verification from appearance- and statistics-based inspection while recognizing intermediate cases.

1 Introduction

  • Sparse real deployment data, rare defects, costly annotations, and changing factory conditions make industrial vision a natural setting for sim-to-real and source-to-target transfer.The relevant mismatch extends beyond rendered-to-camera images to differences in sensors, illumination, materials, production variation, and operational constraints.Source: §1 · p. 2
  • CAD availability is the review’s organizing axis because explicit geometry enables rendering, pose estimation, and test-time visual verification, while CAD-unavailable methods infer abnormality from appearance and learned or statistical priors.Intermediate boundary-prior methods retain weaker correspondence or partial checking through approximate geometry, references, templates, or semantic correspondences.Source: §1 · pp. 2–3
  • The review connects CAD-based detection and pose estimation, represented by T-LESS and BOP [3] [4], with CAD-unavailable anomaly inspection, represented by MVTec AD and VisA.The branches differ in both benchmarks and the priors used to handle deployment variation.Source: §1 · p. 3
  • The review defines industrial visual sim-to-real as a deployment-domain-gap problem, organizes method families by their priors, and uses empirical anchors across detection, pose, and anomaly inspection to examine supported transfer, calibration, and verification.Boundary-prior methods are treated as intermediate cases rather than a separate top-level taxonomy.Source: §1 · p. 4

2 Industrial Vision Tasks and Deployment Domain Gaps

Industrial vision spans diverse recognition, measurement, inspection, and robot-operation tasks, and development conditions often differ from deployment conditions. The review frames sim-to-real around which priors remain available to reduce, calibrate, or verify that mismatch.

Task diversity

  • Industrial vision spans surface-defect and anomaly detection, localization, 6D pose and robot guidance, assembly verification, geometric inspection, and OCR or traceability.Earlier and recent surveys describe applications extending beyond defect classification to inspection, measurement, robot guidance, and process monitoring.Source: §2 · p. 4

Deployment domain gaps

  • Deployment gaps include changes in backgrounds, arrangements, lighting, occlusion, and exposure; pose, calibration, or depth errors can also make robot guidance fail.In inspection, the mismatch also lies between assumptions about abnormality and the real defect mechanisms that emerge in production.Source: §2 · p. 4

Proxy sources

  • Costly or unavailable real data, rare defects, specialized annotation, and limits on production data collection motivate controllable proxies such as CAD rendering, domain randomization [13], synthetic anomalies, normal references, pretrained features, and prompts.CAD rendering varies pose, camera, and illumination, while domain randomization broadens source variation.Source: §2 · p. 4

Limits of source conditions

  • Simulation and proxy sources remain incomplete: rendering approximates real optics, materials, clutter, and sensor artifacts; CAD omits production variation, and synthetic defects may miss physical failure mechanisms.Pretrained features may support coarse recognition without enabling dense localization or calibrated inspection thresholds.Source: §2 · pp. 4–5

Spatial sensitivity

  • Small spatial errors can cause operational failures because industrial decisions often require precise localization, alignment, segmentation, or tolerance checks, even under modest image-level domain shifts.Examples include locating small parts or defects, guiding a gripper, verifying assembly, and avoiding false alarms from normal texture variation.Source: §2 · p. 5

Prior availability

  • Industrial sim-to-real concerns recognition systems built under source conditions that differ from deployment, with available priors shaping which assumptions and test-time evidence can address the mismatch.Priors range from CAD meshes and templates to normal-image memory, synthetic defect generators, pretrained visual backbones, and language prompts.Source: §2 · p. 5

3 Prior Availability as the Organizing Axis

The review organizes industrial visual sim-to-real by which priors remain available at deployment: explicit geometry, partial boundary priors, or appearance and statistical substitutes. These priors change the evidence available for source generation, correspondence, verification, and calibration.

3 Prior Availability as the Organizing Axis

  • A four-channel rubric rates source generation, correspondence, test-time checking, and calibration or decision support as absent, weak, or strong evidence, rather than defining a universal benchmark.Strong CAD-available methods use object-level geometry across these roles; boundary-prior methods retain only part, while CAD-unavailable methods can still have normal-reference or semantic evidence without geometric checking.Source: §3 · p. 6, Table 1 · p. 7

3.1 CAD-Guided Geometry: Rendering and Verification

  • CAD serves both as a pre-deployment renderer and as a deployment-time geometric prior for alignment, pose, segmentation, and consistency checking.Rendering can generate labeled views and support simulation and domain randomization; at deployment, geometry can constrain recognition through render-and-compare, mask, or depth consistency.Source: §3.1 · pp. 7–8

3.2 CAD-Unavailable Inspection: Replacement Priors

  • Without usable object geometry, anomaly inspection relies on normal references, feature deviations, synthetic-anomaly assumptions, or semantic priors rather than geometric render-and-verify.MVTec AD and VisA evaluate anomaly classification or localization from normal training data. Examples include PatchCore [22], EfficientAD [23], DRAEM, SimpleNet, and SuperSimpleNet [24] [25], and WinCLIP and AnomalyDINO [27].Source: §3.2 · p. 8

3.3 Boundary Priors Between CAD-Guided and CAD-Unavailable Regimes

  • Boundary priors—such as approximate meshes, templates, partial scans, reference views, or feature correspondences—preserve limited matching or verification, not full CAD authority.They are classified by which CAD role they retain: source generation, correspondence, calibration, or test-time checking, rather than treated as a third top-level taxonomy.Source: §3.3 · p. 8

3.4 Real Data Across Prior Regimes

  • Real data calibrate synthetic-trained CAD systems, but define deployment normality for CAD-unavailable systems; the review therefore compares prior-conditioned questions, not cross-task leaderboard scores.T-LESS diagnostics, MVTec AD localization, and VisA image-level classification differ in priors, labels, metrics, and deployment constraints, so the review proceeds through CAD-available and CAD-unavailable branches [3–6].Source: §3.4 · pp. 8–9

4 CAD-Guided Sim-to-Real Methods

CAD gives industrial sim-to-real two connected resources: synthetic source data and test-time geometric verification. Neither removes the domain gap; transfer and verification depend on coverage, calibration, and usable deployment evidence.

4 CAD-Guided Sim-to-Real Methods

  • Explicit CAD geometry does not remove gaps in illumination, materials, clutter, depth noise, manufacturing tolerance, or sensor behavior; it enables rendering and test-time geometric checks.The CAD-available literature therefore includes both synthetic-data generation and comparison of real observations against object geometry.Source: §4 · p. 9

4.1 CAD as Source: Synthetic Labels, PBR, and Domain Randomization

  • CAD renderers provide pose-labeled synthetic views for textureless rigid objects, but transfer depends on coverage and calibration rather than image count alone [3] [4].Domain randomization can widen source variation across factors such as lighting, materials, sensor noise, and object context.Source: §4.1 · pp. 10–11

4.2 Model-Based Recognition and Learned 6D Pose

  • Learned 6D pose methods vary in architecture but retain CAD-specific geometry, which lets aligned models support checks unavailable to appearance-only detectors [4] [33] [34] [35] [36].These checks include agreement between rendered silhouettes and observed masks, measured depth, and camera-consistent pose.Source: §4.2 · p. 11

4.3 Render-and-Compare as Test-Time Verification

  • CAD-aware verification combines appearance-based proposals with pose, mask, silhouette, or depth consistency, but depends on reliable inputs and manageable ambiguity.Occlusion, symmetry, reflective or transparent materials, clutter, and latency can weaken or preclude verification.Source: §4.3 · p. 12

4.4 Foundation-Assisted CAD Matching and Segmentation

  • Foundation features help CAD-aware systems match, segment, or represent objects across appearance domains while geometry grounds alignment and verification.This combines transferable features with an object-model prior rather than replacing CAD.Source: §4.4 · pp. 12–13

4.5 Boundary Priors: Approximate Models and Reference-Light Pose

  • Approximate models, reference views, and templates form boundary priors whose support ranges from weak matching or localization to explicit geometric verification.A few views may support matching without full depth or silhouette checks; a semantic prompt provides no object-specific geometry.Source: §4.5 · p. 13

4.6 Synthesis of the CAD-Guided Branch

  • CAD supports transfer as a renderer and test-time prior, but its benefits depend on distributional coverage, calibration, and usable geometric evidence.Foundation features can strengthen geometry-based matching, while proposal quality, calibration, and occlusion handling constrain verification.Source: §4.6 · p. 13

5 CAD-Unavailable Industrial Inspection · 5.1 Normal-Reference Memory and Feature Distributions

CAD-unavailable inspection cannot render, align, or geometrically verify an object, so it relies on replacement priors such as normal images, features, synthetic-defect assumptions, or prompts. This section frames normal-reference appearance as a direct substitute for explicit geometry.

5 CAD-Unavailable Industrial Inspection

  • Without CAD, inspection loses object-specific silhouette and depth verification; its replacement priors supply evidence for appearance-dominated inspection.The review frames the question as what can substitute for explicit geometry amid texture, material variation, and rare defects.Source: §5 · p. 14
  • MVTec AD evaluates anomaly classification and localization from normal training images, while VisA adds varied categories, including challenging PCB and multi-instance scenes.Neither benchmark supplies object meshes for render-and-compare verification; both test whether a normality reference can detect departures under real image variation.Source: §5 · p. 14
  • Table 3 organizes CAD-unavailable inspection families by the prior replacing explicit geometry and the remaining deployment risk.Its comparison is about replacement evidence and risk, not which model wins a benchmark.Source: §5 · p. 14, Table 3 · p. 15

5.1 Normal-Reference Memory and Feature Distributions

  • PaDiM models patch-feature distributions; Deep SVDD and Patch SVDD use one-class objectives; SPADE and PatchCore retain normal-feature memories [22], while DifferNet, FastFlow, and CFLOW-AD model likelihoods with flows.These approaches share the premise that, without object geometry, deployment normality is approximated from real or representative normal appearance.Source: §5.1 · p. 15
  • SSIM-based autoencoder segmentation tests whether an image can be reconstructed as normal; modern feature-memory and flow methods often outperform these simple reconstruction criteria.Both approaches define normality through what can be represented or reconstructed from normal evidence.Source: §5.1 · pp. 15–16
  • Normal real images define the reference distribution, so they must cover legitimate variation in surfaces, materials, lighting, exposure, background, and production batches.The gap emerges when the learned distribution is too narrow or benchmark normality fails to represent deployment normality.Source: §5.1 · p. 16
  • Patch-memory and feature-distribution methods remain strong because normal production data are easier to collect than defects, but usually detect statistical deviation rather than explain defect mechanisms or semantic abnormality.They provide a counterweight to broad claims about foundation models, while leaving assembly logic unexplained.Source: §5.1 · p. 16

5.2 Teacher-Student Residuals and Efficient Inspection

Teacher-student inspection substitutes CAD geometry with residuals from learned normal representations, offering fast, dense anomaly signals while making deployment performance dependent on normal-data coverage and residual calibration.

5.2 Teacher-Student Residuals and Efficient Inspection

  • Teacher-student methods detect abnormal regions through mismatches with learned normal representations, with industrial anchors spanning feature matching, reverse distillation, and millisecond-latency EfficientAD [23].Uninformed Students is an early industrial anchor.Source: §5.2 · p. 16
  • Without CAD, anomaly scores measure residuals against learned normal behavior; these methods can be fast, dense, and trained without defect labels, but deployment variation may trigger false anomalies.Residual reliability depends on the normal training distribution and teacher representation; lighting, material finish, camera settings, or background changes can appear abnormal even when acceptable.Source: §5.2 · p. 16
  • CAD-unavailable inspection needs calibrated masks, controlled false alarms, and thresholds stable under process drift, so evaluation should go beyond image-level AUROC to residual calibration.Strong image-level ranking alone may coexist with unstable dense masks or thresholds.Source: §5.2 · p. 16

5.3 Synthetic-Anomaly Self-Supervision

Synthetic-anomaly self-supervision addresses scarce real defects by synthesizing abnormalities on real normal images or features rather than rendering CAD geometry. Its value depends on whether the synthetic distribution teaches a boundary that transfers to real defect mechanisms.

5.3 Synthetic-Anomaly Self-Supervision

  • Synthetic-anomaly methods learn abnormality by synthesizing defects on real normal images or features, addressing scarce real defects without rendering CAD objects.They form a CAD-unavailable analogue of synthetic-data generation, but their synthetic inputs mimic unknown defects by corrupting normal images or features.Source: §5.3 · pp. 16–17
  • CutPaste, DRAEM [24], Natural Synthetic Anomalies, SimpleNet, SuperSimpleNet, and DeSTSeg use artificial defects or feature perturbations to train anomaly localization and discrimination.Earlier AnoGAN and GANomaly established reconstruction or adversarial deviation as abnormality signals, although they were not designed specifically for modern industrial inspection.Source: §5.3 · pp. 16–17
  • The key test is whether synthetic anomalies teach a decision boundary that transfers to real defect mechanisms, not whether they look plausible in the abstract.Unlike CAD-available recognition, which renders known geometry, CAD-unavailable methods corrupt normal images or features to mimic unknown defects.Source: §5.3 · p. 17
  • Synthetic corruptions may teach useful localization or create shortcuts because real defects include scratches, dents, contamination, missing components, assembly errors, and process-specific failures.Synthetic-anomaly methods are an important source-prior family, not a guaranteed replacement for defect data.Source: §5.3 · p. 17

5.4 Vision-Language Priors for Zero- and Few-Shot Inspection · 5.5 Dense Visual Foundation Features and Few-Shot Reference

These sections distinguish semantic language priors from dense visual-feature priors for CAD-unavailable inspection, while emphasizing that promising zero- and few-shot approaches do not by themselves resolve dense localization or deployment calibration.

5.4 Vision-Language Priors for Zero- and Few-Shot Inspection

  • CLIP’s image-text feature space underlies WinCLIP’s zero- and few-shot anomaly classification and segmentation [27] and later prompt- and scoring-based variants.The later methods explore object-agnostic prompts, challenge-specific pipelines, learnable prompt adaptation, fine-grained descriptions, and mutual scoring of unlabeled images.Source: §5.4 · p. 17
  • Language prompts describing normal and abnormal states may reduce object-specific reference needs and help generalize across categories when new products or defects lack full training sets.In this review, language is a weak prior that partially replaces object geometry and normal-reference memory.Source: §5.4 · p. 17
  • Broad language descriptions may not align with small, texture-, material-, or process-specific defects, so VLMs have not solved dense industrial sim-to-real localization.Deployment requires deciding whether a particular region under a particular camera and material is unacceptable.Source: §5.4 · p. 17

5.5 Dense Visual Foundation Features and Few-Shot Reference

  • DINOv2 provides self-supervised visual representations, adapted by AnomalyDINO for few-shot anomaly detection, while UniVAD reflects training-free or few-shot unified detection.This branch addresses settings without CAD or defect labels but with a few normal references and a strong pretrained visual backbone.Source: §5.5 · pp. 17–18
  • Unlike CLIP’s language-aligned semantics, DINOv2-style patch features preserve spatial structure and texture similarity, making them more like learned normal-reference representations.This may make dense visual features more naturally suited to localization than prompt-driven semantic classification.Source: §5.5 · p. 18
  • Dense foundation features still require calibration of feature maps into reliable image-level decisions and pixel-level masks, including threshold selection, false-alarm control, and production-shift robustness.Appearance comparisons cannot verify object-specific shape or pose without a model or reference structure, unlike CAD-at-test-time geometry.Source: §5.5 · p. 18

5.6 Robustness, Logical Anomalies, and Deployment Realism

The section broadens CAD-unavailable evaluation beyond local defects to logical, structural, 3D, multi-view, and domain-shift scenarios. It argues that deployment decisions require realistic evidence about robustness and operational consequences, while reference-based geometry does not restore full CAD verification.

5.6 Robustness, Logical Anomalies, and Deployment Realism

  • MVTec LOCO AD, MVTec 3D-AD, Real-IAD, RobustAD, and MVTec AD 2 extend anomaly evaluation to logical constraints, geometry, multi-view inspection, domain shifts, and advanced unsupervised scenarios.These datasets broaden the review beyond standard local-texture benchmarks.Source: §5.6 · p. 18
  • Strong performance on curated benchmarks may not transfer to new lighting, camera placement, materials, normal variation, assembly constraints, or multi-view inspection requirements.CAD-unavailable robustness depends on reference coverage, feature stability, calibration, and benchmark realism; CAD geometry can sometimes reject physically inconsistent hypotheses.Source: §5.6 · p. 18
  • Depth and point clouds reveal defects weakly visible in RGB, but without a scaled object model, geometry remains reference-based rather than verified against the whole design.Such boundary methods compare local geometry with normal references, fuse photometric and geometric residuals, or check multi-view consistency.Source: §5.6 · p. 18
  • Deployment requires deciding whether evidence warrants stopping a line, rejecting a product, or requesting human review, with thresholds, false positives, masks, image-level decisions, and explanations all mattering.Robustness datasets and logical-anomaly benchmarks belong in the review even though they are outside its empirical anchor suite.Source: §5.6 · pp. 18–19

5.7 Generative and LVLM Boundary Directions · 5.8 Synthesis of the CAD-Unavailable Branch

Generative and large vision-language models extend CAD-unavailable inspection toward controllable defect synthesis and interactive reasoning, but remain boundary directions. The CAD-unavailable branch is best understood through the appearance-based priors that ground its decisions, not a single cross-task ranking.

5.7 Generative and LVLM Boundary Directions

  • Diffusion and generative methods such as AnomalyDiffusion, SeaS, AnoGen, and GroundingAnomaly synthesize realistic defect images, masks, or grounded anomalies when real defects are scarce.They extend synthetic-anomaly methods toward more realistic and controllable data generation.Source: §5.7 · p. 19
  • AnomalyGPT, IADGPT, IAD-GPT, and AD-Copilot apply large multimodal models to anomaly detection, localization, reasoning, and visual in-context comparison.They may also compare examples, describe defects, support interactive inspection, or provide reasoning for human operators.Source: §5.7 · p. 19
  • Generative and LVLM approaches are boundary and future-work directions, not a third main taxonomy, because they weaken—but do not eliminate—the need for object-specific evidence [91, 98–102].Generated defects must resemble real process failures, and LVLM explanations must rely on reliable visual evidence and calibrated inspection behavior; without CAD, systems still need a substitute for geometric verification.Source: §5.7 · p. 19

5.8 Synthesis of the CAD-Unavailable Branch

  • CAD-unavailable methods define normality or deviation through reference memories, feature distributions, learned residuals, synthetic defects, semantic priors, and transferable foundation features.Robustness benchmarks, generative synthesis, and LVLM methods extend this branch toward more realistic and interactive inspection.Source: §5.8 · p. 19
  • Methods should be compared by their prior: CAD enables alignment and render verification, whereas CAD-unavailable systems test consistency with normal examples, appearance statistics, synthetic defects, or pretrained features.This contrast motivates representative empirical anchors for the two settings rather than one leaderboard across unrelated industrial vision tasks.Source: §5.8 · p. 19

6 Empirical Anchors Across Prior-Availability Regimes

The anchors compare how prior availability shapes industrial sim-to-real: CAD-supported rendering and test-time geometry pose different questions from CAD-unavailable inspection, so their results are not a single leaderboard.

6.1 Experimental Design and Metric Logic

  • The anchors use T-LESS/BOP for CAD-supported detection and pose verification, and MVTec AD and VisA for CAD-unavailable anomaly inspection; their tasks and metrics differ.Renderer-only detection uses mAP50 and mAP50:95, CAD-at-test-time diagnostics use overlap, coverage, recall, and AUROC, and anomaly inspection uses image/pixel AUROC and F1. The anchors are representative rather than exhaustive and do not estimate one effect size across industrial tasks.Source: §6 · p. 20, §6.1 · pp. 20–21

6.2 CAD-Guided Anchor: CAD as Synthetic Renderer

  • mAP50:95 falls from 0.1521 (5k PBR) to 0.1287 (50k), reaches 0.4041 with strong randomization, 0.6265 after 50-image calibration, and 0.7424 with larger capacity plus calibration.Under synthetic domain-randomized training, increasing capacity to YOLOv8s reaches 0.6077 without real object labels; the results indicate that source distribution, detector capacity, and small real calibration contribute differently to transfer.Source: Table 5 · p. 23, Figure 4 · p. 24, §6.2 · p. 25

6.3 CAD-Guided Anchor: CAD at Test Time

  • Oracle CAD verification reaches AUROC 0.9963; practical MegaPose yields 0.7322 full-mask IoU, 0.9218 visible coverage on valid same-class proposals, and depth-confidence fusion improves good-pose AUROC to 0.8804.The oracle geometry check gives 0.8553 mean visible-mask IoU, while the best detector reaches same-class ROI recall@0.5 of 0.8206 at confidence 0.10; practical performance still inherits proposal and pose errors.Source: Figure 6 · p. 27, Table 6 · p. 27, §6.3 · pp. 27–28
  • CAD at test time makes real-image hypotheses geometrically checkable through pose, rendered masks, and depth, adding a verification channel beyond synthetic-data generation.The practical value still depends on detector proposals and pose estimation, as well as occlusion and object ambiguity.Source: §6.3 · pp. 26–28, Figure 6 · p. 27

6.4 CAD-Unavailable Anchor: Normality and Foundation Priors

  • PatchCore gives the best MVTec AD image AUROC and best pixel AUROC on both datasets; EfficientAD-S has stronger MVTec AD pixel F1 and slightly stronger VisA image AUROC.WinCLIP retains moderate MVTec AD image-level signal but much weaker pixel scores, especially on VisA; AnomalyDINO-S approaches PatchCore and EfficientAD-S on MVTec AD, while DINOv2-Large gives the strongest VisA image AUROC without monotonic gains in pixel localization or thresholded masks.Source: Table 7 · p. 29, Figure 7 · p. 30, §6.4 · pp. 30–31
  • CAD-unavailable results vary with category type, and the category heat maps show zero-shot VLM pixel-level limitations across categories rather than only in the average.The section identifies repeated products, texture-like surfaces, small-defect objects, and PCB-style layouts as distinct inspection category types.Source: §6.4 · p. 31

A MVTec AD image AUROC

The normal-reference budget diagnostic finds strong pixel AUROC with sparse normal sets, while the cross-regime synthesis organizes industrial transfer by the prior available for calibration and verification.

6.5 CAD-Unavailable Normal-Reference Budget Diagnostic

  • With test images, anomaly labels, and pixel masks fixed, the diagnostic varies normal-reference coverage across 5%, 10%, 25%, and 100% for PatchCore and AnomalyDINO-S on three categories per dataset.The selected categories cover a low-count object case, structured objects, a hard VisA category, and PCB inspection; 100% rows reuse full-baseline results.Source: §6.5 · p. 33
  • At 5% of normal training images, PatchCore reaches 0.9819 pixel AUROC on selected MVTec AD categories and 0.9712 on selected VisA categories; larger budgets mainly improve image discrimination and thresholded masks.AnomalyDINO-S also remains strong at 5%, especially in image AUROC on VisA.Source: Table 8 · p. 33, §6.5 · p. 34, Figure 9 · p. 34

6.6 Cross-Regime Synthesis

  • Across regimes, deployment performance depends on the available prior: CAD supports rendering, calibration, and test-time pose, rendering, and depth checks, while normal-reference memory and dense foundation features ground appearance-only inspection.The review avoids direct numerical comparison between T-LESS detection metrics and MVTec/VisA anomaly metrics.Source: §6.6 · pp. 34–35
  • CAD availability is not a generic advantage, and boundary priors should be assessed by whether they support rendering, weak matching, semantic localization, or calibration.CAD-unavailable inspection uses normal-reference appearance and pretrained feature spaces rather than simply having less information.Source: §6.6 · p. 35

7 Conclusion and Outlook

Industrial visual sim-to-real is a gap between deployment evidence and the evidence needed for reliable decisions, organized by the prior available rather than by synthetic-to-real transfer alone. Evaluations should make that prior and its evidence channel explicit instead of collapsing detection, pose, and anomaly inspection into one scoreboard.

Conclusion and Outlook

  • CAD availability determines whether geometry can provide labels, calibration, or test-time verification, making it a defining prior rather than a minor dataset attribute.The review frames the deployment gap around available and required evidence, not only synthetic and real images.Source: §7 · p. 35
  • With CAD, source-distribution design, detector capacity, and small real calibration can matter more than render count, while test-time pose, mask, and depth checks enable geometric verification.This channel does not eliminate proposal errors, pose ambiguity, occlusion, or scoring difficulty.Source: §7 · p. 36
  • Without CAD, normal-reference memory and dense foundation features are strong appearance-only substitutes, while zero-shot language-aligned transfer is less reliable for dense localization under standard anomaly protocols.Small normal-reference sets can support pixel-level ranking, but thresholded masks, image-level decisions, and category-specific behavior remain deployment challenges.Source: §7 · p. 36
  • Approximate models, reference views, templates, generated anomalies, prompts, and learned features create a boundary-prior regime, but these evidence types are not interchangeable.The organizing question remains which prior grounds the deployment decision.Source: §7 · p. 36
  • Evaluations should state the available prior and report its relevant evidence and operating conditions, rather than compare detection, pose, and anomaly inspection on a single cross-regime scoreboard.For pre-deployment-only CAD, focus on source-distribution design and real-domain calibration; for test-time CAD, include geometric verification; without CAD, report normal-reference coverage, foundation features, threshold calibration, and category structure.Source: §7 · pp. 36–37
  • Adaptation and uncertainty estimation modify how deployment observations are used, not the prior regime, and should be constrained or monitored against incorrect poses, defect leakage, and threshold drift.Uncertainty is most useful when tied to actions such as accepting, rejecting, requesting another view, lowering line speed, or seeking human review.Source: §7 · p. 36

A Empirical Diagnostics for the Review Anchors

The appendix preserves run-level diagnostics for the review anchors as reproducibility and audit support, not as a new benchmark or standalone experimental comparison. Its evidence is tailored to prior availability: detector training curves and predictions, CAD-at-test-time geometric checks, and CAD-unavailable appearance probes.

Appendix scope

  • The appendix supports reproducibility and audit of completed detector, CAD-at-test-time, and CAD-unavailable probes, without introducing a new benchmark, method family, or separate comparison.It is supporting evidence for the paper’s prior-availability framing, not a standalone experimental section.Source: A Empirical Diagnostics for the Review… · p. 46

Diagnostic formats

  • Because CAD-at-test-time geometry checks and CAD-unavailable memory, feature, or probe workflows lack comparable epoch curves, the appendix uses qualitative contact sheets for those anchors.Epoch-level curves are provided for YOLOv8 runs sharing a supervised training protocol.Source: A Empirical Diagnostics for the Review… · p. 46

CAD-as-renderer diagnostics

  • YOLOv8 [103] diagnostics on T-LESS/BOP [3, 4] show that PBR-only volume does not close the real-domain gap, while randomization, detector capacity, and 5% real calibration improve behavior through distinct contributions.The figures document complementary source-design, capacity, and calibration effects; prediction grids and confusion matrices show the same progression, and label distributions identify the real data as calibration rather than a replacement for synthetic rendering.Source: Figure 10 · p. 47, Figure 11 · p. 48, Figure 12 · p. 49

CAD-at-test-time diagnostics

  • MegaPose [17] overlays on held-out T-LESS [3] scenes show CAD used after detection to geometrically explain observed object layout, not only to generate synthetic images.These additional contact sheets complement Figure 6.Source: Figure 15 · p. 52

CAD-unavailable diagnostics

  • MVTec AD [5] toothbrush examples show ground-truth masks, anomaly heat maps, and predicted masks as an appearance-prior probe for CAD-unavailable inspection.The diagnostic is not directly comparable to detector or pose-estimation training curves.Source: Figure 16 · p. 53

Related papers

Loading 2605.30581v2…