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 lacks a common framework for mismatches between available priors and deployment decisions. This review organizes methods by CAD availability and finds that transfer depends more on source design, detector capacity, calibration, and test-time verification than on synthetic render count alone.

  • Problem

    Industrial vision must transfer across changing sensors, conditions, production variation, and scarce real labels, but existing reviews lack a prior-based framework spanning detection, pose, and inspection.

  • Method

    The review organizes methods into CAD-available, CAD-unavailable, and boundary-prior regimes using a rubric covering source generation, correspondence, test-time checking, and calibration.

  • Results

    Across anchors, synthetic render count alone does not close transfer, while source-distribution design, detector capacity, small real calibration, and CAD-based verification materially shape outcomes.

  • Takeaways & Limitations

    The deployment decision should be grounded in the available prior: CAD enables geometric mask, pose, and depth checking, whereas CAD-unavailable methods rely on calibrated normality and feature deviation.

  • Takeaways & Limitations

    CAD-based test-time verification remains limited by proposals, identity, calibration, depth or masks, pose refinement, occlusion, and difficult materials or clutter.

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

The review frames industrial visual sim-to-real as a deployment-domain-gap problem organized by prior availability, distinguishing CAD-available, boundary-prior, and CAD-unavailable regimes. This axis determines whether deployment decisions can be grounded in explicit geometry or must rely mainly on appearance, statistics, and learned priors.

  • Motivation: Industrial deployment often lacks large labeled real datasets because products, rare defects, and expensive annotations limit available evidence while factory conditions vary.Relevant annotations include pixel-level defect masks, 6D poses, and detailed inspection labels.
  • Problem framing: Industrial sim-to-real includes transfers from CAD renderings, simulated RGB-D, synthetic defects, normal references, pretrained features, benchmark datasets, and language prompts into real factory environments.The review therefore treats sim-to-real more broadly than synthetic rendered images transferring to real camera images.
  • Taxonomy: CAD availability is the primary organizing split because it changes the usable prior, the domain mismatch, and the possible test-time verification mechanisms.CAD or renderable models support synthetic views, pose estimation, and rendered-mask or depth comparison against real observations.
  • Taxonomy: Boundary-prior methods occupy an intermediate regime using approximate geometry, sparse references, templates, category-level shape, foundation-model correspondences, or prompts rather than full CAD.These methods preserve weaker correspondence or partial checking without requiring complete explicit geometry.
  • Literature integration: The review connects CAD-based detection, pose estimation, segmentation, and geometric verification with CAD-unavailable inspection based on normality, feature residuals, synthetic anomaly assumptions, and learned priors.T-LESS and BOP concretize the CAD-available setting with object models, RGB-D observations, and 6D pose annotations for textureless industrial objects.

2 Industrial Vision Tasks and Deployment Domain Gaps

Industrial vision spans recognition, measurement, inspection, robot guidance, and process monitoring, with development conditions often differing from deployment. Industrial visual sim-to-real therefore concerns designing systems that use available priors to reduce, calibrate, or verify deployment mismatch.

  • Task diversity: Industrial vision includes recognition, measurement, inspection, robot guidance, and process-monitoring functions across manufacturing, logistics, quality control, and robotics.The task scope extends beyond generic recognition benchmarks.
  • Deployment gaps: Domain gaps arise when systems trained with synthetic or limited real labels face different backgrounds, object arrangements, lighting, occlusions, or camera exposure at deployment.Detection and localization are practical manufacturing tasks whose development and deployment conditions rarely match.
  • Data constraints: Simulation and proxy data are attractive because real data may be costly, delayed, unavailable, rare, or difficult to annotate for poses, masks, and pixel-level defects.New parts, rare failures, specialized labels, and production constraints can limit balanced real-data collection.
  • Simulation limitations: Simulation remains incomplete because rendering approximates optics, illumination, materials, clutter, noise, blur, and depth artifacts, while models and synthetic defects omit production variability and physical defect mechanisms.The resulting source condition is useful but incomplete for transfer.
  • Precision requirements: Industrial decisions are sensitive to small spatial errors, requiring precise localization, alignment, segmentation, pose guidance, assembly verification, or defect detection with controlled false alarms.These requirements affect cluttered-part detection, robot manipulation, surface inspection, and geometric inspection.
  • Prior-centered framing: Industrial visual sim-to-real is best framed as designing systems for differing source and deployment conditions while using available priors to reduce, calibrate, or verify the mismatch.Relevant priors include CAD meshes, templates, reference views, normal-image memories, synthetic defect generators, pretrained backbones, and language prompts.

3 Prior Availability as the Organizing Axis

Prior availability changes the recognition problem by determining which evidence can support source generation, correspondence, calibration, and test-time verification. The review therefore separates CAD-available and CAD-unavailable branches, treating boundary priors as partial extensions rather than a third top-level regime.

  • CAD-unavailable settings: CAD-unavailable inspection replaces object-level geometry with normal-reference appearance, feature similarity, teacher-student discrepancies, synthetic anomaly assumptions, or thresholded statistical evidence.MVTec AD and VisA represent the common setting of normal training data followed by anomaly classification or localization on held-out real images.
  • Prior-strength rubric: A strong CAD-available method uses object-level geometry for labeled source generation, pixel or pose correspondence, test-time rendering, and calibration or decision support.The operational rubric evaluates four channels: source generation, correspondence, test-time checking, and calibration or decision support.
  • CAD-available settings: CAD has dual roles: pre-deployment rendering for labeled views and domain randomization, and deployment-time geometric support for pose, segmentation, render-and-compare scoring, and depth or mask verification.This makes geometry active during deployment rather than merely a tool for producing synthetic RGB images.
  • Boundary-prior regime: Boundary priors include approximate meshes, templates, partial scans, reference views, and foundation-feature correspondences, supporting limited matching or verification without full geometric authority.They are interpreted by identifying which part of the prior-strength vector they preserve, not as a separate top-level taxonomy.
  • Role of real data: Small real datasets serve different mechanisms: supervised calibration in CAD-available settings and normal-reference modeling in CAD-unavailable settings.Real data can calibrate synthetic-trained detectors or estimate deployment normality, even when real defect labels are unavailable.
  • Review structure: The review rejects a single leaderboard because T-LESS, MVTec AD, and VisA expose different domain-gap responses under different priors, labels, metrics, and deployment constraints.Its two main branches consequently ask how explicit geometry is used before and during deployment, or how normal appearance and anomaly evidence are modeled.

4 CAD-Guided Sim-to-Real Methods

CAD makes industrial sim-to-real geometrically addressable without eliminating appearance, sensing, and deployment gaps. Its benefits depend on distributional coverage and calibration for rendering, and on usable proposals and observations for test-time geometric verification.

  • Synthetic data generation: Explicit CAD geometry supports scalable synthetic data generation, but rendered images remain mismatched with real illumination, materials, clutter, sensing, and manufacturing variation.CAD changes the available responses to the domain gap rather than removing it.
  • Synthetic data generation: Synthetic quantity does not equal coverage: carefully designed variation or small real calibration can transfer better than additional renders repeating the same assumptions.Relevant variation includes lighting, materials, backgrounds, object context, sensor effects, and placement.
  • Test-time geometric verification: Render-and-compare methods such as MegaPose use object models during inference to estimate 6D pose for novel objects.They generate hypotheses and compare rendered evidence with real observations.
  • Test-time geometric verification: At test time, CAD enables pose, mask, silhouette, and depth consistency checks that combine appearance-based proposals with geometric verification.This makes the object model an active inference resource rather than only a training-data source.
  • Test-time geometric verification: CAD verification depends on reasonable proposals, correct identity, calibration, usable depth or masks, and refinable pose hypotheses, while occlusion, symmetry, transparency, reflectance, and clutter weaken it.These conditions are stress tests of the verification channel rather than incidental nuisances.
  • Hybrid and boundary-prior methods: Foundation features and partial geometric priors form intermediate CAD-available regimes, bridging real-appearance variation while retaining geometry for alignment and verification.Boundary priors can include approximate models, partial scans, reference images, category-level shape priors, templates, or foundation representations.

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

CAD-unavailable inspection replaces object-level geometric verification with priors derived from normal examples, feature distributions, synthetic assumptions, pretrained features, or language prompts. Normal-reference methods make real-data coverage central: they can avoid defect examples but must represent legitimate production variation and generally detect deviation without explaining its mechanism.

  • 5 CAD-Unavailable Industrial Inspection: CAD-unavailable systems cannot render targets, align meshes, or verify hypotheses through object-specific silhouette and depth consistency.
  • 5 CAD-Unavailable Industrial Inspection: Replacement priors include normal images, reference examples, category data, synthetic defect assumptions, pretrained visual features, language prompts, and unlabeled test images.
  • 5 CAD-Unavailable Industrial Inspection: MVTec AD evaluates unsupervised anomaly classification and localization from normal images, while VisA adds more varied categories, challenging PCBs, and multi-instance scenes without object meshes.
  • 5 CAD-Unavailable Industrial Inspection: The branch is structured by what evidence replaces explicit geometry and what deployment risk remains, rather than by benchmark model rankings.
  • 5.1 Normal-Reference Memory and Feature Distributions: Normal-reference methods estimate normal appearance in feature space using patch distributions, one-class objectives, stored representative features, or correspondences.
  • 5.1 Normal-Reference Memory and Feature Distributions: Reconstruction methods define normality by whether test images can be reconstructed from normal evidence, while modern feature-memory and flow methods often outperform simple reconstruction criteria.
  • 5.1 Normal-Reference Memory and Feature Distributions: Normal real images define the reference distribution itself, so coverage must include legitimate variation in surfaces, materials, lighting, exposure, backgrounds, and production batches.
  • 5.1 Normal-Reference Memory and Feature Distributions: Patch-memory and feature-distribution methods remain strong because normal production data are easier to collect than defects, but they usually detect statistical deviation without explaining defect mechanisms or semantic abnormality.

5.2 Teacher-Student Residuals and Efficient Inspection

Teacher-student methods substitute learned representation consistency for explicit CAD geometry, detecting anomalies through residuals against normal behavior. Their practical advantages require calibration because residual quality, masks, false alarms, and thresholds depend on the normal training distribution and process stability.

  • Teacher-student methods replace explicit geometry with teacher-student feature consistency learned from normal data, detecting abnormal regions through residuals or mismatches.Uninformed Students, feature-pyramid matching, and multiresolution distillation provide industrial anchors for dense residual signals.
  • Their appeal is practical: teacher-student inspection can be fast, dense, and trained without defect labels, while scoring deviations from learned normal behavior.Unlike CAD-based verification, the method cannot test rendered-object agreement with the image; it evaluates whether representations behave as they did on normal data.
  • CAD-unavailable inspection requires calibrated residual signals, because image-level AUROC alone does not ensure stable masks, controlled false alarms, or thresholds under process drift.Residuals are meaningful only relative to the normal training distribution, and deployment may require calibration beyond feature extraction.

5.3 Synthetic-Anomaly Self-Supervision

Synthetic-anomaly self-supervision addresses scarce real defects in CAD-unavailable inspection by corrupting normal images or features to teach abnormality. Its transfer depends on whether the synthetic-anomaly distribution captures real defect mechanisms, so it is a source-prior family rather than a guaranteed replacement for deployment evidence.

  • Synthetic-anomaly self-supervision: Synthetic-anomaly methods synthesize defects on real normal images or features instead of rendering CAD objects, addressing the scarcity of real defects.AnoGAN and GANomaly established reconstruction or adversarial deviation as abnormality signals, while CutPaste, DRAEM, and related methods use artificial defects or feature perturbations.
  • Synthetic-anomaly self-supervision: The CAD-unavailable analogue of synthetic data generation corrupts normal images or features to mimic unknown defects rather than using known object geometry.The relevant question is whether these corruptions teach a decision boundary that transfers to real defect mechanisms.
  • Synthetic-anomaly self-supervision: Synthetic corruptions may encourage useful localization but also create shortcuts unrelated to deployment defects such as scratches, dents, contamination, assembly errors, or material inconsistencies.Real industrial defects arise from diverse process-specific mechanisms, making visual plausibility alone insufficient.
  • Synthetic-anomaly self-supervision: Synthetic-anomaly methods should therefore be treated as an important source-prior family, not as a guaranteed replacement for real defect mechanisms.Their value depends on how well the synthetic anomaly distribution transfers to deployment conditions.

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

Vision-language models provide weak semantic priors for zero- and few-shot inspection, while dense foundation features provide spatially structured, patch-level alternatives when CAD and defect labels are unavailable. Both remain dependent on industrially reliable localization and calibration rather than resolving dense sim-to-real inspection outright.

  • 5.4 Vision-Language Priors for Zero- and Few-Shot Inspection: Vision-language models introduce a CAD-unavailable prior by aligning images with language, with CLIP and WinCLIP anchoring language-aligned industrial anomaly inspection.Subsequent methods include AnomalyCLIP, APRIL-GAN, AdaCLIP, FiLo, and MuSc.
  • 5.4 Vision-Language Priors for Zero- and Few-Shot Inspection: Language prompts may reduce object-specific references and improve category generalization when new products or defects appear before full training data exists.In this framing, language is a weak prior partially replacing object geometry and normal-reference memory.
  • 5.4 Vision-Language Priors for Zero- and Few-Shot Inspection: Broad semantic prompts do not automatically localize small, texture-level, material-specific, or process-specific defects reliably in industrial images.Deployment requires judging whether a particular region under a particular camera and material is unacceptable.
  • 5.4 Vision-Language Priors for Zero- and Few-Shot Inspection: VLM methods are promising weak-prior approaches, but they have not solved dense industrial sim-to-real inspection.They remain part of the CAD-unavailable regime rather than a replacement for dense industrial verification.
  • 5.5 Dense Visual Foundation Features and Few-Shot Reference: Dense self-supervised features offer a non-language foundation prior, with DINOv2 representations, AnomalyDINO patch features, and UniVAD supporting few-shot or training-free anomaly detection.These methods address settings lacking CAD and defect labels but retaining a few normal references and a pretrained visual backbone.
  • 5.5 Dense Visual Foundation Features and Few-Shot Reference: Unlike CLIP’s language-aligned semantics, DINOv2-style features preserve dense visual correspondence and patch-level spatial structure for texture-level localization.They can function more like learned normal-reference representations than prompt-driven semantic classifiers.
  • 5.5 Dense Visual Foundation Features and Few-Shot Reference: Dense foundation features still require calibration that converts feature distances or maps into reliable image-level decisions and pixel-level masks.Larger backbones and stronger pretraining do not automatically solve threshold selection, false-alarm control, or robustness to production shifts.

5.6 Robustness, Logical Anomalies, and Deployment Realism

This section argues that industrial anomaly detection must be evaluated beyond curated local-defect benchmarks, under logical, structural, geometric, multi-view, and real-world domain variations. Deployment realism also requires deciding whether evidence is sufficient for line stoppage, product rejection, or human review.

  • Benchmark scope: Robustness benchmarks expose failures caused by new lighting, camera placement, materials, normal variation, logical constraints, and multi-view inspection.The review highlights MVTec LOCO AD, MVTec 3D-AD, Real-IAD, and RobustAD as benchmarks expanding beyond standard local-texture anomaly settings.
  • Benchmark scope: MVTec LOCO AD evaluates logical and structural anomalies, while MVTec 3D-AD emphasizes three-dimensional industrial anomaly detection and localization.These benchmarks broaden anomaly evaluation beyond local texture defects and highlight cases where geometry matters without object-level CAD.
  • Three-dimensional inspection: Depth maps and point clouds can reveal dents, missing material, deformation, and assembly errors that are weakly visible in RGB, but they do not restore the full CAD role.Without a scaled object model, depth-based systems can compare local geometry with normal references, fuse photometric and geometric residuals, or check multi-view consistency.
  • Deployment realism: Deployment evaluation must consider thresholds, false positives, mask quality, image-level decisions, and explanation alongside anomaly-score ranking.The operational question is whether evidence is sufficient to stop a line, reject a product, or request human review.
  • Deployment realism: Robustness datasets and logical-anomaly benchmarks belong in the review even when they are outside the current empirical anchor suite.Their inclusion prevents an overly narrow interpretation of CAD-unavailable success.

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

Generative defect synthesis and LVLM inspection define boundary directions for CAD-unavailable industrial inspection, but remain future-work directions rather than a third taxonomy. The CAD-unavailable branch is best synthesized by the role of its appearance, behavioral, synthetic-abnormality, semantic, or pretrained-feature prior rather than by a single cross-task ranking.

  • 5.7 Generative and LVLM Boundary Directions: Generative methods create realistic or controllable defect images, masks, or spatially grounded anomalies when real defects are scarce.Examples include AnomalyDiffusion, SeaS, AnoGen, and GroundingAnomaly.
  • 5.7 Generative and LVLM Boundary Directions: LVLM methods support anomaly detection, localization, reasoning, visual in-context comparison, and defect description for interactive inspection.Examples include AnomalyGPT, IADGPT, IAD-GPT, and AD-Copilot.
  • 5.7 Generative and LVLM Boundary Directions: These generative and LVLM methods are boundary and future-work directions because they weaken object-specific data requirements without resolving what evidence grounds the deployment decision.Generated defects must resemble real process failures, while LVLM explanations must remain tied to reliable visual evidence and calibrated inspection behavior.
  • 5.8 Synthesis of the CAD-Unavailable Branch: Without object-level geometry, CAD-unavailable industrial sim-to-real becomes the definition and calibration of appearance-based priors.The branch includes normal-reference memory, feature distributions, teacher-student residuals, synthetic anomalies, and vision-language methods.
  • 5.8 Synthesis of the CAD-Unavailable Branch: Normal-reference memory and feature distributions estimate deployment normality, whereas teacher-student methods detect residuals from learned normal behavior.Synthetic-anomaly methods substitute for real defects when those defects are rare.
  • 5.8 Synthesis of the CAD-Unavailable Branch: CAD-unavailable families should be compared by prior role: test-image consistency with normal examples, appearance statistics, synthetic-defect assumptions, or pretrained feature spaces.By contrast, explicit geometry in the CAD-available branch enables alignment and render verification, motivating representative empirical anchors rather than a single cross-task ranking.

6 Empirical Anchors Across Prior-Availability Regimes

The empirical anchors compare CAD-available and CAD-unavailable industrial transfer through different datasets, tasks, and metrics rather than a single leaderboard. They show that source-distribution design, detector capacity, calibration, and test-time geometric verification can matter more than synthetic image count, while normal-reference and foundation-feature priors replace CAD in anomaly inspection.

  • Experimental design: T-LESS/BOP evaluates CAD both as a synthetic PBR renderer and as a test-time geometric prior, whereas MVTec AD and VisA evaluate CAD-unavailable anomaly inspection.The branches use different tasks, datasets, and metrics, so their numerical results are not a single leaderboard.
  • CAD-available transfer: 0.1287 mAP50:95 after increasing synthetic PBR images from 5k to 50k contrasts with 0.4041 after strong domain randomization, making source distribution more important than raw image count.The result is interpreted as evidence that CAD rendering is a distribution-design problem rather than a raw image-count problem.
  • CAD-available transfer: 0.6265 mAP50:95 follows from fine-tuning with 50 labeled real images, while YOLOv8s reaches 0.6077 without real object labels and reaches 0.7424 mAP50:95 with the same calibration.Small real calibration and detector capacity provide complementary ways to absorb residual synthetic-to-real mismatch.
  • CAD-at-test-time verification: 0.8553 mean visible-mask IoU from the pose oracle establishes reliable CAD geometry, while MegaPose reaches 0.7322 mean full-mask IoU across proposals and 0.9218 mean ground-truth visible coverage for valid same-class proposals.CAD creates a test-time verification channel by rendering geometry into the real camera view and scoring masks, poses, and depth consistency.
  • CAD-unavailable inspection: PatchCore leads MVTec AD image AUROC and pixel AUROC on both datasets, while AnomalyDINO-S approaches reference methods on MVTec AD and DINOv2-Large gives the strongest VisA image AUROC.WinCLIP retains moderate image-level signal but performs substantially worse for dense pixel localization, especially on VisA; category structure also affects CAD-unavailable inspection.

A MVTec AD image AUROC

On selected MVTec AD and VisA categories, CAD-unavailable methods retain strong pixel-level AUROC with small normal-reference sets, while broader coverage mainly improves image discrimination and thresholded masks. The diagnostic shows that real images serve as normality reference evidence rather than calibration for a rendered source domain.

  • Interpretation: In CAD-unavailable settings, normal-reference images provide deployment evidence for calibrated normality, complementing dense visual foundation features as appearance-only alternatives.The review contrasts this evidence source with CAD used as a renderer, calibration scaffold, or test-time geometric prior.
  • Budget diagnostic: The budget diagnostic subsamples only normal training images while keeping test images, anomaly labels, and pixel masks unchanged.PatchCore and AnomalyDINO-S use 5%, 10%, 25%, and 100% of normal training images across three selected MVTec AD and three selected VisA categories.
  • Results: Pixel AUROC remains high with small normal sets, whereas image-level discrimination and thresholded F1 improve more gradually as reference coverage increases.Because the test set is fixed, the budget changes only how much normal appearance evidence the method can store or compare against.
  • Results: 0.9819 pixel AUROC is achieved by PatchCore with 5% of normal images on selected MVTec AD categories, and 0.9712 on selected VisA categories.AnomalyDINO-S also remains strong at 5%, especially for image AUROC on VisA.

7 Conclusion and Outlook

The review frames industrial visual sim-to-real around prior availability: CAD enables geometry-aware transfer and test-time verification, while CAD-unavailable and boundary-prior regimes rely on calibrated appearance, feature, semantic, or intermediate evidence. It recommends making the prior, evidence channel, and deployment operating point explicit in evaluation and reporting.

  • Conclusion and Outlook: CAD availability determines whether industrial sim-to-real can use explicit geometry for source-data generation, calibration, pose estimation, segmentation, and deployment verification.The review treats CAD as a defining prior rather than a minor dataset attribute.
  • Conclusion and Outlook: Render count alone does not determine transfer quality; source-distribution design, detector capacity, and small amounts of real calibration can matter more.With CAD available at test time, pose, rendering, mask overlap, and depth consistency provide a geometric verification channel.
  • Conclusion and Outlook: CAD-unavailable methods replace explicit geometry with normal-reference appearance, feature distributions, teacher-student residuals, synthetic-anomaly assumptions, semantic prompts, or dense foundation features.Normal-reference memory and dense visual foundation features are described as strong appearance-only substitutes, while zero-shot language-aligned transfer is less reliable for dense localization.
  • Conclusion and Outlook: Approximate models, reference-light pose methods, CAD-template segmentation, generative anomaly synthesis, and large vision-language systems define a boundary-prior regime with intermediate evidence.This regime weakens strict assumptions about what is known before deployment without eliminating prior availability as the organizing question.
  • Conclusion and Outlook: Future evaluations should state the available prior before selecting methods, datasets, or metrics, then match evaluation channels to predeployment, test-time, or absent-CAD evidence.The review calls for source-distribution and calibration evaluation when CAD is predeployment-only, geometric verification when CAD persists at test time, and normal-reference, feature, and threshold reporting when CAD is absent.
  • Conclusion and Outlook: Test-time adaptation, self-training, and uncertainty estimation modify evidence use after the prior is specified rather than defining separate prior regimes.CAD-guided adaptation should remain constrained by geometric consistency to avoid reinforcing incorrect poses.

A Empirical Diagnostics for the Review Anchors

The appendix preserves run-level diagnostics supporting the empirical anchors and prior-availability framing. It shows that CAD-rendered transfer depends on source design, detector capacity, and limited real calibration, while CAD-at-test-time and CAD-unavailable branches require different diagnostics.

  • Appendix purpose: The appendix provides reproducibility and audit evidence for completed detector, CAD-at-test-time, and CAD-unavailable anchor outputs.It does not introduce a new benchmark, method family, or separate comparison.
  • Branch-specific diagnostics: CAD-at-test-time and CAD-unavailable anchors use qualitative diagnostics because their inference-time geometry checks and memory, feature, or probe workflows lack comparable epoch curves.MegaPose overlays visualize CAD as a geometric explanation after detection, while synthetic-anomaly examples probe appearance-prior behavior.
  • CAD-as-renderer diagnostics: YOLOv8 training diagnostics show that PBR-only synthetic volume does not by itself close the real-domain gap.The runs compare synthetic volume, domain randomization, larger detector capacity, and small real calibration on T-LESS/BOP.
  • CAD-as-renderer diagnostics: Randomized sources, larger capacity, and small real calibration progressively improve real-image detection behavior across prediction grids and class-level confusion matrices.These diagnostics show stronger real-domain behavior and diagonal structure than PBR-only models.
  • CAD-as-renderer diagnostics: The small-real rows use real data as calibration evidence rather than replacing synthetic rendering.B3 and B6 use a much smaller 5% real calibration set alongside synthetic data.
Loading 2605.30581v2…