Deep Reading Notes: Comparative Evaluation of 3D Reconstruction Methods for Object Pose Estimation

1. Basic information

• Short name: Reconstruction → Pose Benchmark
• Authors: Varun Burde, Assia Benbihi, Pavel Burget, Torsten Sattler
• Affiliation: Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague
• Venue: WACV 2025
• Published paper: CVF Open Access
• Preprint: arXiv:2408.08234
• Data and code: reconstruction_pose_benchmark
• Reading mode: Deep Dive

2. Main assessment

This paper reframes reconstruction quality as asset quality within 6D pose systems. Multiple methods reconstruct each YCB-V object, replacing CAD in MegaPose / FoundationPose under fixed test images and BOP metrics. Completeness, accuracy, PSNR, and visual appeal do not reliably predict pose AR; downstream evaluation reveals whether useful geometry and appearance cues survive.

Five conclusions matter more than any single reconstruction ranking:

1. CAD replacement works for some objects: reconstructions of simple, large, Lambertian objects sometimes approach CAD oracles.
2. An overall gap remains: the best reconstructions typically retain about a 10%–20% pose quality gap from CAD.
3. Classical MVS remains strong: RealityCapture is fast, stable with sparse views, and competitive with strong neural methods; Nerfacto also offers a favorable accuracy–time tradeoff.
4. Assets need not be visually perfect: incomplete meshes or incorrect textures can support good poses if discriminative parts remain.
5. Specific properties dominate difficulty: small, reflective, finely textured objects and appearance differences from legacy test objects are particularly challenging.

The CAD-free boundary matters: new captures generate deployment assets, but YCB-V CAD still supports reconstruction evaluation, registration, and oracles. Each reconstruction needs manual coarse ICP initialization. This tests replacing CAD during pose inference, not a fully CAD-free onboarding workflow.

2.1 Guide to the paper’s figures and tables

Figure / table	Main variables	Causal boundaries
Figure 1	Geometry metrics, MegaPose / FoundationPose AR, runtime, PSNR	Related metrics are not equivalent; ranking conflicts reject a single reconstruction metric as a pose proxy
Figure 2	Uniform, small Lambertian, and shiny groups	Size, materials, fine texture, and product versions matter; textureless does not always mean difficult
Figure 3	Image count, pose AR, completeness, incomplete meshes	The authors identify a useful 75–100-image budget, not universal saturation; discriminative parts matter more than completeness

3. Research question and task boundaries

3.1 Why general pose estimators still depend on CAD

FoundationPose, MegaPose, and related methods generalize to unseen objects without per-test-object retraining, but render-and-compare, template generation, or correspondences usually need accurate 3D models. Industrial CAD, laser scans, and high-quality manual meshes are expensive; service robots cannot prepare CAD for every possible object.

Capturing RGB images around new objects is more realistic. The question therefore goes beyond visual resemblance:

$$\text{reconstructed asset}\quad\xrightarrow{\text{fixed pose estimator}}\quad \text{object pose accuracy}$$

Can it approach:

$$\text{CAD asset}\quad\xrightarrow{\text{same pose estimator}}\quad \text{oracle pose accuracy}?$$

3.2 Controlling benchmark variables

The intended variable is reconstruction method, not pose estimator. Main controls include:

• Fixed 21 YCB-V objects and test frames;
• Fixed MegaPose / FoundationPose checkpoints and inference;
• All reconstructions registered to the same YCB-V CAD frame;
• Method-specific camera pose refinement disabled in favor of shared refined camera extrinsics;
• CAD baseline as oracle;
• Shared BOP evaluation protocol.

Residual uncontrolled variables include:

• RealityCapture on RTX 3060 / Windows versus mostly one A100 for other methods;
• Different preprocessing, representations, meshing, and texturing paths;
• Officially recommended look-alike replacements for discontinued YCB objects;
• Manual ICP initialization;
• Unequal computation paths for implicit and mesh rendering.

3.3 What pose task does the main set represent?

• Instance-level: 21 specific YCB-V objects;
• Asset available at inference: a reconstruction or CAD for every object;
• Generalizable estimator: the pose network is not trained separately for each of these 21 objects;
• Full 6D pose evaluation: BOP VSD / MSSD / MSPD rather than orientation alone;
• RGB / RGB-D routes: MegaPose uses RGB; FoundationPose uses RGB-D;
• Assets are the study subject, not estimator architecture.

4. Building the benchmark data

4.1 Why recapture instead of reusing YCB-V videos?

YCB-V test images deliberately challenge pose estimation with clutter, occlusion, complex backgrounds, and limited target pixels. These hinder complete, high-quality reconstruction. Reconstructing directly from test videos would confound unsuitable inputs with algorithm weaknesses, preventing fair comparison of reconstruction methods.

The benchmark therefore separates:

1. Reconstruction component: newly captured, object-focused, unobstructed, high-resolution calibrated views;
2. Pose component: established YCB-V / BOP test images and GT poses.

The question becomes whether assets reconstructed from favorable inputs can handle difficult downstream pose scenes.

4.2 Capture hardware and trajectory

Item	Setting
Camera	Basler Ace acA2440-20gc
Robot arm	7-DoF KUKA LBR IIWA 14 R820
Camera mounting	Robot flange
Image resolution	2448×2048
Latitude	0°–150°, 10° intervals
Longitude	0°–360°, 10° intervals
Images per object	397–505, averaging 480
Exposure	Fixed during scanning
Content	Full object within the field of view, unobstructed and object-focused
Releases	Calibrated / undistorted images, object masks, 3D reconstructions

Camera positions follow an evenly distributed spherical trajectory. IIWA stack, MoveIt, and CHOMP handle planning and collision avoidance.

4.3 Calibration and extrinsic refinement

OpenCV and MoveIt handle camera / hand-eye calibration, with reported errors up to 5 mm translation and 0.1° rotation. Robot kinematic extrinsics remain sensitive to encoder noise, especially near degenerate extended configurations.

To share camera poses across reconstruction methods, two offline refinement stages are used:

1. iNGP pose refinement;
2. COLMAP bundle adjustment.

Extrinsics are then fixed and method-specific pose refinement disabled, avoiding camera quality differences when one method optimizes cameras and geometry while another accepts fixed cameras.

4.4 Look-alike replacements and domain mismatch

Discontinued YCB-V products are replaced with recommended look-alikes that may differ from old CAD / test images in texture and scale. Mean size change is 4% of object dimensions. The captured physical scale is preserved rather than forced to match old CAD.

This has two implications:

• Practical relevance: packaging changes routinely create mismatches between legacy images and new objects;
• Causal confounding: pose drops can reflect product version changes as well as reconstruction algorithms.

Legacy / updated object grouping partly separates these effects, but users should still record object versions as metadata.

5. Benchmark Pipeline

5.1 Reconstruction

Inputs include calibrated multiview RGB, intrinsics, refined extrinsics, and masks. Possible outputs include:

• textured mesh;
• colored mesh;
• Meshes extracted from SDF / density fields;
• Direct novel-view rendering from implicit representations.

Meshes are preferred where possible for consistent coordinates and efficient downstream rendering. Direct implicit rendering is an additional comparison that can preserve view-dependent appearance but is usually expensive.

5.2 Mesh Registration

YCB-V GT uses original CAD coordinates, while reconstructions use robot / world coordinates. A rigid transform is needed to reuse GT. The procedure is:

1. Manually initialize coarse alignment;
2. Run ICP between reconstruction and YCB-V CAD;
3. Crop residual background after alignment.

This is a substantial implicit manual dependency. Fully CAD-free deployment would need another canonicalization / reference registration scheme; this paper does not solve it.

5.3 Pose Inference

• Replace the estimator’s CAD with each reconstruction;
• Run MegaPose on YCB-V RGB tests;
• Run FoundationPose on RGB-D tests;
• Run the same estimator with CAD for each object’s oracle;
• Evaluate in BOP format.

The benchmark measures relative degradation: AR lost by substituting reconstruction for CAD in the same estimator, and stability of reconstruction rankings.

6. Compared reconstruction methods

6.1 Classical MVS

COLMAP

SfM / MVS produces dense point clouds, followed by Poisson surface reconstruction. Default reconstruction parameters are used and room backgrounds roughly cropped. When COLMAP’s texturing path is insufficient, colored meshes outperform external MVS texturing for pose; colored meshes are used in the main results.

RealityCapture

This commercial MVS system emphasizes efficiency. Running on RTX 3060 / Windows, RealityCapture has among the shortest reconstruction times and stable sparse-view performance, competing with strong learned methods under pose-oriented evaluation.

6.2 Surface / SDF Methods

Method	Representative design	Representation used for pose
UniSURF	Occupancy network approximating a surface	Extracted mesh
NeuS	Neural SDF + volume rendering	Marching Cubes mesh
VolSDF	SDF-constrained density	Mesh
MonoSDF	Monocular geometry / depth cues constrain the SDF	Mesh
BakedSDF	VolSDF-based mesh baking with vertex view-dependent appearance	Baked mesh
NeuralAngelo	Multiresolution hash grids for high-resolution surfaces	Mesh

6.3 Radiance / Grid Methods

Method	Representative design	Engineering consideration
iNGP	Hash-grid-accelerated NeRF	Extracted / colored meshes and direct implicit rendering are both evaluated
Nerfacto	Nerfstudio integration of NeRF improvements including adaptive sampling	Strong pose performance, but slow downstream direct rendering
Plenoxels	Explicit feature / density grids + spherical harmonics without an MLP	Faster training than conventional NeRF
SplatFacto	Nerfstudio splatting approach	Weaker pose performance in Figure 1

The paper reports that evaluating all methods with just one pose estimator takes roughly 8 A100 GPUs running for 8 days. This motivates representative coverage rather than claims of exhaustive reconstruction benchmarking.

6.4 Texturing and direct rendering

• Learned mesh face colors are obtained by querying the networks;
• RealityCapture uses proprietary texturing;
• Main COLMAP / iNGP results use colored meshes because off-the-shelf MVS texturing hurts pose;
• iNGP also supports direct network rendering;
• Direct Nerfacto rendering within MegaPose takes nearly 2 minutes per pose image on one A100, making high-quality novel-view synthesis potentially unsuitable for large candidate-rendering budgets.

The interface between reconstruction representation and pose renderer is itself a research variable. Imperfect meshes do not rule out useful implicit rendering, while attractive PSNR does not guarantee pose-discriminative render-and-compare features.

7. Evaluation metrics

7.1 Geometry Completeness

Sample reference CAD surface points and find reconstructed nearest neighbors. Completeness is the fraction below a distance threshold: how much true surface is covered.

7.2 Geometry Accuracy

Sample reconstructed surface points and find nearest reference CAD points. Accuracy is the fraction below threshold: how much reconstructed surface lies near the reference.

Curves use 1 mm–2 cm thresholds. To reduce density-dependent nearest-neighbor statistics, each reconstruction is first subsampled to a fixed count.

7.3 Appearance PSNR

Rendered appearance PSNR measures mesh appearance, but averages pixel errors without knowing which textures / edges disambiguate poses or matching the estimator’s learned feature space.

7.4 BOP Pose Metrics

VSD

Visible Surface Discrepancy compares visible surfaces under GT / predicted poses. It handles symmetry, occlusion, and visibility; the paper relates it to global positioning / navigation.

MSSD

Maximum Symmetry-Aware Surface Distance measures the worst surface distance over equivalent symmetries. Small offsets near high-curvature grasp regions make it informative for manipulation risks.

MSPD

Maximum Symmetry-Aware Projection Distance measures pixel projection misalignment, relevant to AR / VR rendering alignment.

Recall is computed over error thresholds and averaged over thresholds and test images into each metric’s AR, then the three ARs are averaged into overall BOP AR.

These are not simple scalar translation / rotation errors: they apply poses to object models and assess symmetry-aware surface / projection misalignment.

8. Main results: why task-aware evaluation is necessary

8.1 Overall CAD oracle gap

The CAD oracle lies at high AR on Figure 1’s far right; even strong reconstructions generally fall below it. The text summarizes a roughly 10%–20% gap. Adequate geometry in many cases does not mean CAD can universally be removed.

VSD recall is closer to baseline, suggesting broadly preserved positioning. Larger MSSD / MSPD drops reveal manipulation surface offsets and visual overlay misalignment respectively.

8.2 Counterexamples for geometric metrics

The evidence includes explicit ranking conflicts, not just correlations:

• RealityCapture and BakedSDF have similar completeness / accuracy, but BakedSDF gives slightly better poses;
• Nerfacto has worse geometry than MonoSDF but similar poses;
• VolSDF and MonoSDF geometric curves nearly overlap, while pose evaluation distinguishes them;
• Lower UniSURF / iNGP geometry agrees with lower pose performance.

Geometry provides a necessary foundation without a reliable monotonic mapping to pose. Error locations matter more than mean coverage: missing a broad plane but preserving distinctive edges can be better than complete geometry with discriminative details smoothed away.

8.3 Counterexamples for appearance metrics

• RealityCapture has higher PSNR than BakedSDF but worse poses;
• COLMAP and NeuralAngelo have similar PSNR but different poses;
• iNGP direct implicit rendering outperforms iNGP mesh rendering for pose.

PSNR weights pixels equally, while estimators may emphasize contours, local textures, and synthetic–real feature consistency. High PSNR can reflect background or broad low-frequency colors without preserving discriminative cues.

8.4 Ranking agreement across two estimators

Figure 1 includes MegaPose and FoundationPose: despite RGB versus RGB-D inputs, top reconstruction rankings broadly agree. The authors suggest reconstruction improvements may therefore benefit pose estimators generally.

This site’s more cautious interpretation:

• Agreement shows the findings are not purely a MegaPose artifact;
• Both methods share CAD / render-and-compare principles, limiting extrapolation to keypoints, coordinate regression, or direct pose networks;
• Universality would require more estimators with different representations and training approaches.

8.5 Accuracy–runtime tradeoff

RealityCapture and Nerfacto occupy favorable AR–runtime regions in Figure 1. RealityCapture runs on a weaker RTX 3060 yet is much faster than many single-A100 neural methods. BakedSDF has strong pose quality but substantially higher time costs.

For fast pose asset creation, advanced neural rendering need not be the best system choice. Put capture, reconstruction, meshing, pose runtime, and final AR on one Pareto plot.

9. Object property analysis

The 21 objects are divided into 8 groups by shape, texture, material, and changes from original physical objects. Figure 2 shows 3 groups; others appear in the supplement.

9.1 Large, Lambertian, uniformly textured objects

These sometimes approach the CAD oracle. This does not establish that texture is universally unimportant; rather, for these objects:

• Geometry is relatively easy to reconstruct;
• Large contours and surfaces sufficiently constrain pose;
• Texture adds little information, limiting the harm from texturing imperfections.

Stable geometry can therefore matter more than photorealistic texture for pose, challenging the intuition that textureless objects always make the hardest reconstruction assets.

9.2 Small objects (≤8 cm)

At fixed image resolution, small objects have fewer usable texels and lower reconstructed texture resolution. For approximately symmetric geometry, fine texture may distinguish orientation; blur / baking errors can remove it and cause symmetry confusion.

9.3 Shiny / reflective objects

Reflections vary with viewpoint and resist static UV textures. Although NeRF methods can represent view-dependent effects, Figure 2 shows no consistent dominance over RealityCapture. The authors identify room for improving texture generation and downstream interfaces.

9.4 Legacy vs updated objects

Updated packaging or materials can mismatch look-alike reconstruction textures with older YCB-V test objects. Better poses on legacy objects indicate appearance domain mismatch as a separate asset risk from geometry.

9.5 No universal winner

No top method consistently dominates on the hardest object types. Neither “NeRF solves reflections” nor “RealityCapture suits everything” follows. A more reasonable system estimates object properties before choosing reconstruction / rendering pipelines.

10. Image counts and sparse-view reconstruction

10.1 Data budget curves

Uniform spherical subsets contain 25, 50, 75, 100, 150, 300, and all images. The authors describe 75–100 views as enough for near-best reconstructions; some neural pose AR curves in Figure 3 still improve substantially afterward. Treat this as an empirical budget for these captures, not universal saturation.

Curves are not strictly monotonic:

• Some neural methods fail to converge reliably with 25 / 50 views;
• Robot encoder noise depends on configuration rather than being i.i.d.;
• Adding noisy camera poses may temporarily harm geometry;
• Pose AR and 5 mm completeness do not change in lockstep.

Selecting well-covered, well-calibrated, informative views is more accurate advice than assuming every added image helps.

10.2 RealityCapture efficiency with sparse views

Figure 3’s caption reports:

• 25 images: under 1 min on average;
• 50 images: about 2 min on average;
• Relatively stable pose performance as image counts decrease.

RealityCapture is thus a strong practical baseline. Neural reconstruction claims for onboarding should at least compare its Pareto curves at 25 / 50 / 100 views.

10.3 Poor appearance need not mean poor poses

Exact Figure 3 examples:

• Two reconstructed bowls both miss surfaces; one matches CAD oracle AR=38, while the other reaches only AR=15;
• Two wood blocks reach AR=39, exceeding CAD oracle AR=27, even though one has visibly incorrect texture.

The explanation is that reconstructed discriminative parts such as outer bowl edges can support matching. A more complete mesh can still fail if it loses key cues.

A reconstruction exceeding its CAD oracle does not prove greater reconstruction accuracy. Possible explanations include:

• Textures better matching test appearance;
• CAD material rendering gaps from real images;
• Estimator biases that happen to favor particular artifacts;
• Per-object sampling variation.

Do not extrapolate isolated AR=39 > 27 examples into general reconstruction superiority over CAD.

11. Are the claims supported?

Claim	Evidence	Strength	This site’s assessment
Standard reconstruction metrics cannot reliably predict pose	Figure 1 ranking conflicts	Strong	Supports the benchmark’s core claim
Classical methods compete with neural reconstruction	RealityCapture AR / runtime and sparse-view curves	Strong	Holds for this YCB-V pipeline
Top reconstruction rankings agree across two estimators	Figure 1 center	Moderate–strong	Supports two render-and-compare estimators, not every method
Reconstructions can replace CAD for some objects	Figure 2 groups approaching oracles	Moderate–strong	Supported for simple / large / Lambertian groups, not universally
Best reconstructions retain a substantial CAD gap	Figure 1 and reported 10%–20%	Strong	Especially clear for MSSD / MSPD and hard objects
The authors identify 75–100 views as a useful budget	Figure 3 curves	Moderate	Empirical for these controlled captures; some methods keep improving, without a shared saturation point

12. Critical analysis

12.1 Strengths

1. Appropriate downstream objective: Chamfer / completeness / PSNR are not assumed to equal pose utility.
2. Broad variables: geometry, appearance, poses, runtime, image counts, and object properties.
3. Strong fixed estimators: MegaPose + FoundationPose span RGB and RGB-D.
4. Established YCB-V / BOP tests: avoids inventing easier pose tests and supports community reuse.
5. Open data: calibrated images, masks, reconstructions, and code enable new methods.
6. Honest engineering conclusions: classical MVS, manual registration, unequal hardware, and sparse-view failures are visible.

12.2 Limitations

1. CAD registration dependency: automatic canonical coordinate definition for new objects remains unresolved.
2. Manual coarse alignment: costly at scale and potentially operator-dependent.
3. Only 21 objects: YCB lacks transparent, large industrial, thin sheet, and strongly repetitive structures.
4. Favorable captures: unobstructed full objects, fixed exposure, and known camera poses differ from online robotic reconstruction amid clutter.
5. Look-alike confounds: geometry, updated textures, and scale differences jointly affect poses.
6. Unequal hardware / software: runtime is not pure algorithmic complexity.
7. Similar estimator paradigms: both use render-and-compare, limiting extrapolation to other families.
8. Limited statistical intervals: main plots average results without per-method uncertainty / repeated captures.
9. Detection is outside the analysis: model-based pose evaluation omits detector / segmentation failures from end-to-end success.

12.3 Implicit assumptions

• Near-complete unobstructed spherical capture is possible;
• Robot kinematics and extrinsics can be corrected offline;
• Object identity is known;
• Reference CAD is available for registration / evaluation;
• Objects remain stationary during capture;
• Asset and test object versions agree or differ acceptably;
• Estimators accept a shared mesh / rendering interface;
• Object masks are available;
• Reconstruction scale is preserved or correctly aligned.

12.4 Three common misreadings

”Geometry does not matter” is incorrect

High pose AR often accompanies good geometry, while poor geometry limits pose. One geometric summary metric cannot reliably rank task utility; geometry still matters.

”75 images suffice for every method” is incorrect

This is the authors’ empirical summary for controlled spherical capture; some plotted methods still improve with more images. View distributions, extrinsics, materials, and reconstruction families all affect the required budget.

”CAD can already be removed” is incorrect

CAD still supports benchmark registration, and difficult objects retain a 10%–20% gap. Reconstructions can replace CAD in some deployment settings, with the benchmark helping identify those conditions.

13. Reproduction and extension checklist

13.1 Data capture

• Record camera, lens, resolution, exposure, and white balance;
• Save raw / undistorted images;
• Report at least latitude / longitude coverage histograms;
• Save synchronized robot joint states and flange poses;
• Supply an object mask for every image;
• Record legacy / updated product versions, physical dimensions, and materials;
• Record additional lighting details for transparent / reflective objects.

13.2 Camera Calibration

• Report intrinsic reprojection error separately;
• Report hand-eye translation / rotation errors;
• Save extrinsics before and after refinement;
• Use shared fixed refined cameras for all reconstruction methods;
• Apply controlled camera pose perturbations rather than reporting only the best final calibration.

13.3 Reconstruction standardization

• Fix foreground crops and masks;
• Record input views, iterations, and GPU hours per method;
• Separate training / optimization, meshing, texturing, and registration times;
• Standardize mesh units and scale;
• Fix surface sample counts before geometric metrics;
• Keep direct implicit rendering and extracted mesh results separate.

13.4 Registration

• Save manual coarse alignments;
• Report ICP residuals and failures;
• Test multiple initializations;
• Add CAD-free canonical registration for claims of fully CAD-free onboarding;
• Evaluate registration error sensitivity of pose AR.

13.5 Pose Evaluation

• Fix MegaPose / FoundationPose versions, checkpoints, detectors, and renderers;
• Use identical test detections for CAD oracles and every reconstruction;
• Report VSD / MSSD / MSPD individually, not only overall AR;
• Report per-object, per-property-group, and aggregate results;
• Separate pose runtime from asset construction time;
• Check systematic failures from mesh coordinate / scale errors.

13.6 Statistics

• Repeat image-count sampling and report mean ± std;
• Bootstrap capture trajectories;
• Compute Spearman / Kendall correlations while retaining ranking counterexamples;
• Regress against object size, reflectance, symmetry, and texture frequency jointly;
• Treat updated-object domain mismatch as a separate covariate.

14. Improvement directions

14.1 Pose-aware reconstruction objectives

Current reconstruction optimizes photometric / SDF / geometric losses. Backpropagating fixed-estimator feature matching or render-and-compare scores into assets could prioritize discriminative edges, symmetry-breaking textures, and grasp-critical surfaces.

This risks estimator overfitting. Multiple estimators or a task set—BOP AR, rendering fidelity, collision geometry, grasp contact error—would be more robust.

14.2 Automatic pipeline selection

Estimate size, reflectance, texture frequency, symmetry, and viewpoint coverage from a few views, then select:

• Lambertian / textured objects: RealityCapture or COLMAP;
• view-dependent appearance: NeRF / 3DGS direct rendering;
• geometry-critical manipulation: SDF / depth-assisted reconstruction;
• very small object: macro imaging / higher resolution / active view planning.

14.3 Active View Planning

Figure 3 shows image count is not the only factor. Reconstruction uncertainty and pose utility could guide the next view toward missing discriminative parts instead of uniform capture of 480 images.

The objective could be:

$$v^*=\arg\max_v\; \mathbb E[\Delta\operatorname{AR}_{pose}\mid v],$$

rather than maximizing global completeness.

14.4 CAD-free Canonicalization

Full CAD removal requires:

• Metric scale for new reconstructions;
• canonical orientation;
• object origin;
• Transforms to reference images / downstream robot coordinates;
• symmetry group.

Possible tools include fiducial-free robot calibration, known support planes, gravity, category canonical priors, reference pose graphs, or multiview foundation models, with registration error included in evaluation.

14.5 3DGS / neural renderer interfaces

iNGP direct rendering can outperform extracted meshes for pose, suggesting meshing loses useful appearance. Future comparisons could include:

• 3D Gaussian Splatting direct renderer;
• distilled fast neural renderer;
• mesh geometry + learned neural texture;
• hybrid SDF surface + view-dependent appearance;
• Rendering in pose feature space rather than RGB.

Candidate rendering speed is critical: nearly 2 minutes per test image is unsuitable for many pose hypotheses even at good AR.

14.6 Asset-level failure prediction

Predict whether a reconstruction is pose-ready from geometry maps, novel-view consistency, rendered feature gaps, object properties, and coverage, outputting downstream task risks:

• navigation-ready;
• AR-ready;
• grasp-ready;
• Not ready / additional captures needed.

This matches the evidence better than one completeness threshold.

15. Lessons for this 6D Pose website and research

1. Model assets separately: CAD, reconstruction, textures, and registration constrain pose performance alongside estimators.
2. CAD-free is not binary: separate CAD-free training, CAD-free inference, CAD-free registration, and reference-image-only requirements. This paper addresses only part of that spectrum.
3. Choose metrics by downstream tasks: VSD, MSSD, and MSPD differ in navigation, manipulation, and AR relevance; overall AR alone hides this.
4. Engineering benchmarks need runtime and image budgets: asset onboarding is part of deployment cost.
5. Failures teach more than attractive meshes: incomplete yet pose-useful meshes motivate attention to discriminative geometry over visual completeness.

16. Open questions

Data questions

• Are 21 objects enough to train reliable pose-readiness predictors?
• How much does look-alike texture / size mismatch contribute per object?
• How much would reconstruction from original YCB-V videos degrade performance?

Method questions

• Does direct neural rendering’s high pose AR come from view-dependent appearance or estimator preferences for neural artifacts?
• What are the separate effects of mesh simplification, UV atlases, and texture resolution on AR?
• Can a feature-space asset metric better approximate pose utility than PSNR?

Protocol questions

• How long is manual alignment, and how variable are operators?
• How do ICP registration errors of 1–10 mm / 1–5° affect BOP AR?
• Are FoundationPose / MegaPose detections and initializations fully matched?
• Do methods fail on the same objects, or do aggregate rankings hide complementarity?

17. Overall assessment

This benchmark replaces the unreliable inference “better-looking reconstruction means better poses” with reproducible downstream measurement through fixed estimators. Figure 1 ranking conflicts, Figure 2 object groups, and Figure 3 image-count / failure cases show pose utility depends jointly on geometry, appearance, representation, object properties, and renderer interfaces rather than one reconstruction score.

CAD is not declared obsolete. The study identifies when reconstructions approach CAD, where 10%–20% gaps remain, and why small, shiny, finely textured objects and sparse views are bottlenecks. For CAD-free pose, NeRF / 3DGS assets, robotic onboarding, or synthetic rendering research, this is a stricter experimental model than meshes plus Chamfer / PSNR alone.