Deep Reading Notes: CosyPose: Consistent Multi-view Multi-object 6D Pose Estimation

1. Basic information

• Title: CosyPose: Consistent Multi-view Multi-object 6D Pose Estimation
• Authors / affiliations: Yann Labbé, Justin Carpentier, Mathieu Aubry, Josef Sivic; ENS / CNRS / PSL, INRIA, École des Ponts / Univ Gustave Eiffel / CNRS, CIIRC CTU
• Year / conference or journal: 2020 / ECCV
• Paper: arXiv:2008.08465
• Code / project page: Project / Code
• Reading mode: Deep Dive

2. Positioning in one sentence

CosyPose is an RGB-only, instance-level, seen-object system for multi-view, multi-object 6D pose estimation. It generates 6D object candidates per view, estimates unknown camera relationships through symmetry-aware object matching, and jointly optimizes camera and object poses through object-level bundle adjustment to obtain a consistent 3D scene reconstruction.

2.1 Guide to the paper’s figures and tables

Figure / table	Role in the paper	Reading guidance
Figure 1	Inputs / outputs: multiple RGB images → complete scene model	Establish the task boundary; CosyPose is more than a single-frame pose estimator
Figure 2	Method overview: candidate generation → matching → scene refinement	Refer back to it throughout the method
Figure 3	Qualitative candidates and final reconstruction	Read alongside Tables 2 / 3 to understand why multiple views help
Table 1	Single-view pose-estimator results and ablations	Check whether the Stage 1 front end is strong enough
Table 2	Main multi-view results	Supports the central claim that multi-view consistency improves results
Table 3	ADD-S error before / after refinement	Direct evidence for Stage 3 object-level BA
Figure 16	Failure cases: consistent mistakes	Essential for limitations: consistency constraints can consistently preserve errors too

3. Problem setting and task boundaries

• Task definition:
  • Given multiple RGB images of the same static scene, estimate 6D poses of multiple known object instances and the camera pose of each view.
  • The paper defines the problem as recovering a compact scene model: object counts, classes, object poses, and camera poses.
• Inputs:
  • Multiple RGB images.
  • Camera intrinsics for each image.
  • 3D / CAD models corresponding to known object labels.
  • A 2D detector supplies boxes and classes for the single-view stage.
• Outputs:
  • Class and 6D pose of each physical object instance.
  • Pose of each camera / view.
  • Per-view object-pose predictions obtained by projecting the global scene into each image.
• Training inputs / outputs:
  • Training the single-view pose estimator requires real or synthetic RGB crops, object models, and ground-truth poses.
  • The network takes a real crop + rendered crop and predicts a pose update.
  • Multi-view matching and scene refinement are not learned and require no training.
• Inference inputs / outputs:
  • Inputs are multiple RGB images, camera intrinsics, CAD models, and a detector / pose estimator.
  • Output is one consistent scene rather than independent per-image predictions.
• CAD / depth / masks / camera intrinsics / multiple views:
  • CAD / 3D models: required.
  • Depth: not required.
  • Masks: no explicit masks in the main pipeline, but 2D detection is required.
  • Camera intrinsics: required.
  • Camera extrinsics: not required as input; estimated by the method.
  • Renderer: required for single-view render-and-compare and projection during optimization.
  • Multiple views: required for the core setting; the single-view pose estimator can also operate independently.
• Key assumptions:
  • The scene is static, with input views observing the same set of objects.
  • Object candidates have sufficient cross-view consistency; camera-pose validation generally needs at least three inlier object-candidate pairs.
  • Objects belong to a known database with reliable 3D models and symmetry annotations / handling.
• Fairness of baseline settings:
  • Single-view YCB-Video comparisons use the same PoseCNN detections / coarse estimates as DeepIM, making the setting relatively fair.
  • Single-view T-LESS comparisons use the same RetinaNet detections as Pix2Pose, helping isolate pose-estimator differences.
  • In multi-view comparisons, CosyPose does not use ground-truth camera poses, whereas MCN uses known camera poses; the authors emphasize better results with fewer priors.
  • My analysis: directly comparable baselines are scarce for this relatively new setting. The claim relies jointly on single-view ablations, MCN comparisons, and experiments varying view count.

4. Research context

• Is this a pioneering work in a subfield?
  • It did not originate 6D pose or multi-view pose estimation.
  • It is an early representative system that jointly handles RGB-only input, unknown camera extrinsics, multiple views, and multiple objects.
• Main foundations:
  • The single-view refiner clearly builds on DeepIM’s render-and-compare / iterative-refinement approach.
  • The multi-view stage adopts robust matching and global optimization from SfM / bundle adjustment, replacing local points with object-level pose hypotheses.
• Related method designs or paper structures:
  • DeepIM: single-object pose refinement.
  • MCN / A Unified Framework for Multi-view Multi-class Object Pose Estimation: multi-view pose aggregation, but with known camera poses and stronger restrictions.
  • Object-level SLAM: joint object–camera state estimation, often dependent on depth or temporal continuity.
• Main differences from prior work:
  • Does not directly depend on local-feature matching for camera extrinsics.
  • Requires neither a depth sensor nor continuous-motion assumptions.
  • Handles multiple instances of the same class and automatically recovers object counts.
  • Uses object-pose hypotheses as cross-view geometric constraints instead of first recovering point clouds or keypoint tracks.
• Subsequent developments:
  • The project page and repository show subsequent use in the BOP Challenge 2020 and models for multiple datasets.
  • Whether later work comprehensively exceeds it under the same setting still needs verification. Later multi-view pose work may move toward feature-metric alignment, foundation-model detection, and category-level / novel-object settings.

5. Research motivation

• Problems addressed:
  • Ambiguity, false positives, missed detections, and pose noise in single-view RGB 6D pose.
  • Unknown camera poses and difficult object association across views.
  • Scene-level consistency under symmetry, weak texture, multiple instances, and occlusion.
• Reduced assumptions / dependencies:
  • Known camera extrinsics are unnecessary, unlike many multi-view methods.
  • No depth is required, unlike RGB-D object-level SLAM.
  • Views need not be consecutive video frames; temporal continuity is unnecessary.
• Application value:
  • Robots can capture multiple RGB images with moving or multiple cameras for more stable object poses and scene explanations.
  • Particularly relevant to textureless industrial objects / T-LESS-style objects.
• Core approach:
  • Accept that single-view networks produce noisy / incorrect hypotheses.
  • Use cross-view object-level geometric consistency to filter errors and aggregate correct candidates.
  • Globally optimize all objects and cameras in a common coordinate frame.

6. Method details

6.1 Overall pipeline

• Stage 1: object candidate generation
  • Run the detector on each image to obtain 2D boxes / labels.
  • Run the single-view pose estimator for each detection, producing $T_{C_a O_{a,\alpha}}$.
  • Each candidate contains a view index, candidate index, label, pose, and score.
• Stage 2: object candidate matching
  • Run RANSAC for every view pair.
  • Sample object-candidate pairs with the same label to estimate relative camera pose.
  • Identify inlier pairs using a symmetry-aware distance.
  • Build a graph of all pairwise matches; connected components represent physical object instances.
• Stage 3: scene refinement
  • Initialize camera and object poses in a shared world frame.
  • Minimize object-level reprojection error.
  • Jointly optimize all physical objects and cameras with Levenberg–Marquardt.

6.2 Key Modules

Module	Input	Output	Role	Interface to the next module
2D detector	RGB image	Object boxes / labels / scores	Supplies crops and classes to the single-view pose estimator	Each detection generates a pose candidate
Single-view pose estimator	Real crop, rendered crop, object model, initial pose	Refined 6D pose	Generates object-pose hypotheses per view	Stage 2 treats them as object-level measurements
Symmetry-aware distance	Object label, two poses, model points, symmetry set	Pose distance	Handles equivalent poses of symmetric objects	Used in candidate matching, loss, and refinement error
2-view RANSAC matching	Candidates from two views	Relative camera pose, inlier candidate pairs	Estimates view relationships and filters inconsistent candidates	Inlier edges from view pairs enter a graph
Scene-level graph matching	Pairwise inlier edges	Connected components / physical objects	Groups candidates across views into physical objects	Each component becomes one object to optimize
Object-level bundle adjustment	Physical-object groups, camera intrinsics, candidate poses, CAD points	Global camera and object poses	Jointly optimizes one consistent scene	Produces the final multi-object scene reconstruction

6.3 Underlying Models and Methods

• CNN / detector:
  • The paper mentions detectors such as Faster R-CNN and RetinaNet. Experiments use RetinaNet for T-LESS and PoseCNN detections / coarse estimates for YCB-Video.
• Pose refiner:
  • Inspired by DeepIM, but uses EfficientNet-B3 and removes auxiliary flow / mask predictions.
  • Takes a real crop and synthetic rendered crop, then predicts a pose update.
• Rotation representation:
  • Uses a 6D rotation representation: two predicted 3D vectors are orthogonalized into a rotation matrix.
• Geometric solver:
  • Stage 2 uses RANSAC.
  • Stage 3 uses Levenberg–Marquardt optimization.
• Does it use newer methods such as DINO, CLIP, SAM, NeRF, 3DGS, or VGGT?
  • This 2020 paper does not use these foundation-model / neural-rendering methods.

6.4 Training and supervision

• Is it training-free?
  • The overall system is not training-free: single-view detection and pose estimation require training.
  • Multi-view matching / BA are geometric optimization without training.
• Training data:
  • Real training images supplied by YCB-Video and T-LESS.
  • Approximately one million additional synthetic images rendered for each dataset.
  • CAD models for T-LESS and reconstructed models for YCB-Video.
• Ground-truth / pseudo-ground-truth sources:
  • Dataset annotations supply real-data ground-truth poses.
  • Rendering supplies synthetic-data ground-truth poses.
• Loss function:
  • The authors propose a symmetric disentangled loss.
  • Its core idea separates translation, relative depth, and rotation effects, with symmetry-aware distances handling equivalent poses.
  • The symmetry-aware distance can be written conceptually as:

$$D_l(T_1,T_2)=\min_{S\in S(l)}\frac{1}{|\mathcal{X}_l|} \sum_{\mathbf{x}\in\mathcal{X}_l}\|T_1S\mathbf{x}-T_2\mathbf{x}\|_2$$

• Alignment between training objectives and final metrics:
  • The loss is closely related to ADD-S / symmetry-aware pose error, but is not identical to BOP VSD / mAP.
  • My analysis: training targets single-view pose quality; final multi-view quality also depends on detector recall, matching consistency, and BA initialization.

6.5 Inference and engineering complexity

• Inference steps:
  1. Run the detector on each image.
  2. Run the coarse / refinement pose network on each detection.
  3. Perform object-level RANSAC for each view pair.
  4. Build a graph, remove isolated vertices, and form physical objects.
  5. Initialize global camera / object poses.
  6. Perform object-level BA.
  7. Optionally apply 3D NMS to remove duplicate physical objects.
• Test-time optimization / refinement:
  • Required: both single-view pose refinement and final scene refinement are test-time computations.
• Speed / memory / dependencies:
  • The paper reports approximately 320 ms for a common configuration with 4 views and 6 detections per view.
  • Stage 1 takes approximately 190 ms, Stage 2 approximately 40 ms, and Stage 3 approximately 90 ms.
  • Requires PyTorch, a renderer, BOP toolkit, CAD models, and dataset-specific detectors / models.
• Reproduction challenges:
  • The environment is old; the original repository states that it is no longer maintained and points users with issues to Simple-Robotics/cosypose.
  • Multi-dataset BOP formatting, CAD-model paths, rendering setup, and pretrained-weight downloads constitute the main engineering costs.
  • Reproduction requires more than running the network: detection, candidate filtering, RANSAC thresholds, and BA initialization must match the paper.

7. Experimental analysis

7.1 Datasets

• Datasets used:
  • YCB-Video: 92 scenes with approximately 1000 images each; 80 training and 12 test scenes; 21 YCB objects with noise, lighting variation, and occlusion.
  • T-LESS: 20 scenes and 30 textureless industrial objects, mostly symmetric and visually similar. Images captured around scenes provide richer multi-view observations.
• Custom dataset construction and scale:
  • No new benchmark is introduced, but the authors generate many synthetic training images.
  • Synthetic generation: randomly sample 3-9 objects in a 50 cm box, half floating and half placed through physics simulation; cameras face the box center, with background / color domain randomization and augmentation.

7.2 Evaluation Metrics

• Metrics:
  • ADD, ADD-S, ADD(-S).
  • AUC of ADD-S / ADD(-S).
  • VSD recall.
  • ADD-S 0.1d, ADD(-S) 0.1d.
  • mAP@ADD-S0.1d.
• Interpretation:
  • ADD: mean model-point distance under predicted and ground-truth poses, mainly for asymmetric objects.
  • ADD-S: nearest-neighbor / closest-point distances reduce penalties for equivalent poses of symmetric objects.
  • VSD: accounts for visible-surface differences, making it appropriate for T-LESS symmetry / visibility issues.
  • mAP@ADD-S0.1d: combines pose correctness with detection ranking, closer to object-detection precision–recall evaluation.
• Do these metrics sufficiently reflect practical performance?
  • Authors’ claim: the metrics demonstrate improved single-view and multi-view pose accuracy.
  • My analysis: they reflect benchmark accuracy, while robotics also requires grasp tolerances, real-time performance, transparent / reflective-object handling, dynamic scenes, camera-motion control, and failure recovery.

7.3 Compared Methods

Method	Type	Input setting	Main characteristics	Relationship to this paper
PoseCNN	Single-view 6D pose	RGB-D / RGB detections + pose	Representative YCB-Video baseline	CosyPose uses its YCB detections / coarse estimates for fair comparison
DeepIM	Single-view refinement	RGB + initial pose + CAD render	Iterative render-and-compare refinement	Direct design basis of the CosyPose single-view refiner
PVNet	Keypoint voting + PnP	RGB + CAD keypoints	Single-view keypoint method	Single-view YCB baseline
Pix2Pose	Dense correspondence	RGB	RGB-only method on T-LESS	Single-view T-LESS baseline
Implicit	Implicit orientation learning	RGB	RGB-only method on T-LESS	Single-view T-LESS baseline
MCN / Unified multi-view framework	Multi-view pose	Multiple views + known camera poses	Multi-view aggregation	Main multi-view baseline; CosyPose does not use ground-truth camera poses
Object-level SLAM	Scene-level object / camera estimation	Usually RGB-D / temporal	Joint object–camera state optimization	Similar goal, with more sensor or consecutive-frame requirements

7.4 Main results

• State of the art at publication:
  • Authors’ claim: substantially exceeds the contemporary state of the art in both single-view and multi-view 6D object-pose estimation.
  • The project page also identifies CosyPose as the BOP Challenge 2020 winner.
• Improvements over contemporary methods:
  • Single-view YCB-Video: AUC of ADD-S 89.8 versus DeepIM 88.1; AUC of ADD(-S) 84.5 versus DeepIM 81.9.
  • T-LESS SiSo: VSD recall 63.8 versus Pix2Pose 29.5 and Implicit 26.8.
  • Multi-view YCB-Video: 5-view AUC of ADD-S 93.4 versus MCN 80.2.
  • Multi-view T-LESS: with 1, 4, and 8 views, AUC of ADD-S rises from 72.1 to 76.0 to 78.9; mAP@ADD-S0.1d rises from 55.0 to 61.6 to 69.0.
• Gap to the current best under the same setting, if needed:
  • Requires further verification. This is a 2020 method; the latest results for multi-view, RGB-only, known-object estimation with unknown extrinsics need a separate check of current BOP / arXiv / benchmarks.

7.4.1 Main Results Reorganized

Source table	Dataset / task	Method or view count	Metric	Result	Interpretation
Table 1(a)	Single-view YCB-Video	PoseCNN	AUC of ADD(-S)	61.3	Earlier RGB-D / detection baseline
Table 1(a)	YCB-Video single-view	PVNet	AUC of ADD(-S)	73.4	keypoint + PnP baseline
Table 1(a)	Single-view YCB-Video	DeepIM	AUC of ADD-S / ADD(-S)	88.1 / 81.9	Directly related render-and-compare baseline
Table 1(a)	Single-view YCB-Video	CosyPose	AUC of ADD-S / ADD(-S)	89.8 / 84.5	Stage 1 is already a strong front end
Table 1(b)	T-LESS SiSo	Implicit / Pix2Pose	VSD recall	26.8 / 29.5	Older methods struggle with texturelessness and symmetry
Table 1(b)	T-LESS SiSo	CosyPose	VSD recall	63.8	Substantial contributions from single-view refinement and training
Table 2(a)	Multi-view YCB-Video	MCN, 5 views	AUC of ADD-S	80.2	Multi-view baseline using known camera poses
Table 2(a)	Multi-view YCB-Video	CosyPose, 5 views	AUC of ADD-S	93.4	Substantially higher without ground-truth camera poses
Table 2(b)	T-LESS ViVo	CosyPose, 1 / 4 / 8 views	mAP@ADD-S0.1d	55.0 / 61.6 / 69.0	More views stabilize overall detection and pose
Table 3	YCB / T-LESS	Before / after refinement	ADD-S error mm	6.40 → 5.05 / 4.43 → 3.19	Stage 3 clearly reduces inlier-candidate pose error

7.5 Ablation studies

Ablation	Claim tested	Result	Sufficiency
T-LESS w/o loss	Symmetric disentangled loss helps	63.8 falls to 60.1	Broadly sufficient, but tested only in this setting
T-LESS w/o network	EfficientNet / the new architecture improves on the earlier design	63.8 falls to 59.5	Supports this component’s value
T-LESS w/o rotation parametrization	The 6D rotation representation is more stable	63.8 falls to 61.0	Supportive, with a moderate gain
T-LESS w/o data augmentation	Augmentation is crucial for the domain gap	63.8 falls to 37.0	Strong evidence of a substantial synthetic-to-real gap
Before / after scene refinement	Object-level BA improves accuracy	YCB 6.40 mm falls to 5.05 mm; T-LESS 4.43 mm falls to 3.19 mm	Supports the core optimization module
Number of views	Multiple views improve scene-level pose	T-LESS metrics increase over 1/4/8 views	Supports multi-view consistency
Camera pose vs. COLMAP / ground-truth camera poses	Object hypotheses recover camera poses	COLMAP succeeds on only 4% of T-LESS 8-view cases, versus approximately 74% for CosyPose; ground-truth camera poses improve results only slightly	Supports the importance of the unknown-extrinsics setting

7.6 Robustness and Failure Cases

• Occlusion:
  • Multiple views compensate for missed detections in some views; the final scene can recover an object if it is visible and matched elsewhere.
  • An object correctly detected in only one view may be discarded without a cross-view match.
• Textureless / weakly textured objects:
  • T-LESS specifically covers textureless industrial objects, and CosyPose performs strongly on it.
  • Its advantage comes from object-level hypotheses rather than local-feature matching.
• Symmetric objects:
  • Explicit object symmetries enter matching and distance measures.
  • Continuous axial symmetry is approximated in practice with 64 discrete rotation angles.
• Depth noise / mask errors / domain gap:
  • No depth is used, so there is no depth-noise dependency.
  • Masks are not directly required, but a 2D detector is.
  • Domain gaps remain; data augmentation is a crucial T-LESS ablation.
• Failure cases:
  • Consistent mistakes: incorrect candidates consistent across at least two similar views enter the final reconstruction.
  • Missing final objects: correct candidates appearing in only one view cannot be validated across views and are removed.
  • Incorrect camera pose: too few inliers can cause camera-pose estimation to fail.
  • Duplicate objects: visually similar objects can generate overlapping physical-object hypotheses, requiring 3D NMS.

7.7 Are the Contributions Experimentally Supported?

Claimed contribution	Corresponding experiment	Sufficiency	My assessment
Strong single-view refiner	YCB-Video / T-LESS single-view results + ablations	Fairly strong	Many technical changes, but experiments support overall effectiveness
Multi-view object matching recovers consistent scenes	YCB-Video 5 views, T-LESS 1/4/8 views	Fairly strong	Few baselines, but a clear trend
Known camera poses are unnecessary	MCN / COLMAP / ground-truth-pose comparisons	Persuasive	One of the paper’s most valuable setting differences
Scene refinement improves pose accuracy	Before / after ADD-S error	Sufficient	Approximately 20% relative improvement directly supports the claim
Robust to occlusion, false positives, distractors, and symmetry	Qualitative appendix + T-LESS results	Moderate	Extensive qualitative evidence; quantitative breakdowns could be stronger

8. Critical analysis

Strengths

• Naturally combines multi-view geometry with learned single-view pose estimation.
• Clear task boundary: known objects + multi-view RGB + unknown camera extrinsics, a common robotics problem.
• Symmetry is handled throughout training losses, matching distances, and BA.
• Multi-view consistency improves accuracy while filtering false positives, distractors, and single-view errors.
• Public code and models; the project page says the repository includes the full method, synthetic-data generation, and BOP models.

Weaknesses

• Still requires known objects / CAD; it does not solve novel-object or category-level generalization.
• Detection and single-view pose estimation are bottlenecks; missed detections and persistent false positives affect the final scene.
• Multi-view matching needs enough candidates and becomes unstable with few objects or insufficient view information.
• The multi-module pipeline is harder to reproduce than end-to-end methods.
• The paper does not address more complex scenes involving transparency, strong reflections, non-rigid objects, or moving objects.

Implicit Assumptions

• All views come from the same static scene.
• CAD-model scales are accurate and camera intrinsics reliable.
• Object symmetries are known or can be discretely approximated.
• Enough objects are visible in common across views to establish object-level correspondences.
• The 2D detector provides reasonable recall; otherwise the back end has no candidates to recover.

Practical Deployment Feasibility

• Can this support real 6D object-pose estimation or related tasks?
  • Yes, especially tabletop robotics, known industrial parts, and active observation with multiple or moving cameras.
  • RGB-only input and unknown extrinsics lower deployment requirements relative to RGB-D / calibrated multi-camera methods.
• What remains for practical deployment?
  • More stable detectors / segmentors.
  • Temporal consistency in online video.
  • Guided matching / re-detection for candidates appearing only once.
  • More modern rendering, feature matching, or foundation-model front ends.
  • Maintained code environments and data-download scripts.

9. Future Work and Improvement Directions

• Future work mentioned in the paper:
  • Exploit temporal continuity to accelerate toward real-time operation.
  • Grow candidates for missing objects by reprojecting detections from other views, analogous to guided matching.
  • Address consistent mistakes with more views or more required associated candidates.
  • Incorporate object-pose estimation into an active visual-perception loop.
• Directions I would explore:
  • Replace RetinaNet / PoseCNN detections with stronger detectors / segmentors, such as a foundation detector + instance segmentation.
  • Replace pure 2D reprojection loss with feature-metric or differentiable-rendering refinement.
  • Extend object-level BA to joint category-level shape / pose optimization.
  • Explicitly incorporate uncertainty into multi-view matching instead of hard-threshold RANSAC.
  • Add temporal tracking to reuse the previous scene graph under continuous moving-camera observations.
• Modules to replace or strengthen:
  • Detector: Mask R-CNN / SAM-like segmentor / open-vocabulary detector.
  • Pose refiner: modern render-and-compare network, diffusion pose hypothesis refinement, dense feature alignment.
  • Matching: graph neural networks or probabilistic data association.
  • Optimization: robust loss + uncertainty weighting + differentiable renderer.
• Additional experiments:
  • Stability curves versus object count, shared visibility, and camera baseline.
  • Failure analysis as detector recall decreases.
  • Engineering-cost / accuracy tradeoffs against calibrated-camera and RGB-D settings.
  • Real-robot grasp success rather than pose metrics alone.

10. Implications for my research

• Directly reusable ideas:
  • Treat per-view poses as noisy object-level measurements and integrate them through global geometric consistency.
  • Go beyond independent per-object prediction by explicitly recovering scene graphs / physical-object components.
  • Handle symmetry throughout losses, matching, and evaluation rather than correcting only at the end with ADD-S.
• Reusable code / modules / data processing:
  • Multi-view object-candidate matching.
  • object-level bundle adjustment.
  • synthetic data generation pipeline.
  • BOP-format data preparation and evaluation logic.
• Essential baselines:
  • CosyPose is a core baseline for known CAD + multi-view RGB + multi-object pose research.
  • For single-view pose refinement, its refiner represents a major successor to DeepIM.
• Potential research entry points:
  • Reduce missed / incorrect detections with a stronger front end while retaining CosyPose’s scene-consistency back end.
  • Extend object-level consistency to novel objects / category-level pose.
  • Add active view selection based on current scene uncertainty.
  • Use uncertainty-aware optimization for pose-hypothesis confidence and detector scores.

11. Open questions

• Conceptual questions:
  • Does each matching-graph connected component always represent one physical object? How does the graph split or merge when similar objects are close together?
  • Object-level BA uses pose-induced reprojection without directly using image appearance; could this limit final accuracy?
• Formula questions:
  • How are truncated L2 thresholds set in Eqs. 5 / 6, and how sensitive is outlier robustness to them?
  • Are 64 angles sufficient to discretize a continuous symmetry axis, and can this affect high-precision optimization?
• Experimental questions:
  • T-LESS multi-view ViVo lacks a strong direct baseline; how should later methods be compared fairly?
  • When benchmark metrics do not penalize certain false positives, is mAP@ADD-S0.1d a sufficient supplement?
  • The gap between CosyPose-related models and modern methods on the current BOP leaderboard needs further verification.
• Reproduction questions:
  • The original repository is no longer maintained; environment dependencies, pretrained-model downloads, and renderer configuration need particular attention.
  • Can the multi-view back end be reproduced independently by supplying candidates from another method?
  • With a modern segmentation / foundation-model detector, how should candidate scores enter the original RANSAC + graph + 3D NMS pipeline?