Deep Reading Notes: Deep Learning-Based Object Pose Estimation: A Comprehensive Survey

1. Basic information

• Title: Deep Learning-Based Object Pose Estimation: A Comprehensive Survey
• Authors: Jian Liu, Wei Sun, Hui Yang, Zhiweng Zeng, Chongpei Liu, Jin Zheng, Xingyu Liu, Hossein Rahmani, Nicu Sebe, Ajmal Mian
• Year / journal: 2026, International Journal of Computer Vision (IJCV)
• DOI: 10.1007/s11263-025-02646-6
• Paper: arXiv / Open author-manuscript PDF
• Project page: Awesome Object Pose Estimation
• Reading mode: Deep Dive

2. Positioning in one sentence

The survey organizes deep-learning pose estimation in a two-dimensional cross-classification of task formulations against input modalities, training paradigms, inference modes, and benchmarks / metrics. For this site, its evaluation chapter provides the most valuable hierarchy, from rotation angles to BOP AR and from 9DoF IoU to shape CD.

2.1 Guide to the paper’s figures and tables

Figure / page	What to read	Destination in the metric library
Page 5, Section 2.4	Formulas (1)–(10) and the 3DoF / 6DoF / 9DoF classification	Rotation Error, ADD, ADD-S, BOP, 3D IoU, and CD detail pages
Table 1	Instance-level ADD(S)-0.1d and ADD-S AUC	Distinguish error functions from success / AUC reporting protocols
Category-level method tables	5°5cm and 3D IoU	Threshold and box metrics for category-level 9DoF
Novel-object method tables	Mixed reporting of BOP AR, ADD(-S), and others	Metric aliases and paper reverse-lookup rules

3. Problem setting and task boundaries

• Task: recover object rotation and translation from images, depth, point clouds, reference observations, or object models, with size / shape when required.
• formulation:
  • Instance-level: training / test object instances are usually known, often with CAD models.
  • Category-level: test objects may be unseen instances within known categories, requiring shape / size variation handling.
  • Novel-object: instances or categories are unseen during training, relying on reference images, CAD, video onboarding, or generalizable representations.
• Outputs: 3DoF rotation, 6DoF rigid pose, or 9DoF pose + size; articulated objects add joint variables.
• Key boundary: shape-reconstruction CD is auxiliary evaluation and cannot replace pose accuracy.

4. Research context

• Rather than classifying only by architecture, the survey includes segmentation, correspondence, template matching, voting, pose solving / regression, and refinement as inference modes.
• This explains why two papers both labeled 6D pose may be incomparable: depth, CAD, seen-object status, supplied masks / detectors, and metrics all change task difficulty.
• This site’s dataset and evaluation pages should present dataset constraints beside metric protocols, rather than only providing an overall leaderboard.

5. Research motivation

• The authors attribute pose-estimation difficulty both to occlusion, texturelessness, transparency / reflection, and domain gaps, and to task-dependent definitions of correct pose.
• The same prediction has different risks for grasping, AR overlays, and category-level size estimation: some tasks need strict 3D surface alignment, while others emphasize 2D projection or visible surfaces.
• Evaluation therefore receives its own chapter rather than an appendix to dataset descriptions.

6. Evaluation Framework in Detail

6.1 From Error Functions to Final Scores

Predicted pose / shape
        │
        ├── Error function: Rotation / Translation / ADD / VSD / MSSD / MSPD / IoU / CD
        │
        ├── Success rule: < 0.1d, < 5px, 5°5cm, BOP threshold
        │
        └── Aggregation: mean / recall / AUC / AR / AP / mAP

This chain is the most important checklist when reading results. Listing ADD-AUC, BOP AR, and 5°5cm alongside ADD / VSD conflates error measurement with scoring.

6.2 3DoF: Geodesic rotation error

The survey measures orientation differences through the geodesic distance between two rotation matrices:

$$e_R=\arccos\left(\frac{\operatorname{tr}(R_{gt}^{\mathsf T}\hat R)-1}{2}\right).$$

• Authors’ statement / fact: a common 3DoF pose metric, combinable with classification correctness and angular thresholds to obtain accuracy.
• My analysis: it needs no CAD and suits category-level tasks, but equivalent rotations of symmetric objects will be penalized unless minimized over the symmetry group first.

6.3 6DoF: BOP, ADD, Thresholds, and Image Projection

Level	Metric	What it measures	Common misreading
Visible surface	VSD	Compares observable surfaces using rendered depth, observed depth, and visibility masks	Treating it as complete-mesh alignment
Worst-case 3D	MSSD	Maximum vertex displacement after minimizing over the global symmetry set	Treating it as an averaged ADD variant
Perceptual projection	MSPD	Maximum pixel-projection displacement after minimizing over symmetries	Assuming small pixel errors guarantee correct 3D position
Mean model distance	ADD	Mean 3D distance of corresponding CAD points	Assuming it remains fair for symmetric objects
Closest-point distance	ADD-S	Mean nearest-neighbor distance from ground-truth points to the predicted point set	Assuming every texture-defined orientation may be ignored
Joint success	n° mcm	Whether both rotation and translation fall below thresholds	Omitting thresholds or units
Image projection	2D Projection	Mean pixel displacement of corresponding CAD points	Directly comparing 5px across resolutions

6.4 BOP AR and the Updated Detection AP Protocol

The survey describes BOP-M as Average Recall aggregated over VSD, MSSD, and MSPD. The official BOP localization definition can be written as:

$$AR_D=\frac{AR_{VSD}+AR_{MSSD}+AR_{MSPD}}{3}.$$

• Each $AR_e$ averages recall across multiple correctness thresholds; VSD additionally involves multiple misalignment tolerances $\tau$.
• AR therefore needs a subscript or task scope: AR_VSD, AR_D, and AR_C mean different things.
• Protocol update included in these notes: BOP 2024 6D detection uses precision–recall / AP based on MSSD and MSPD, omitting VSD from detection scoring. An older BOP AR summary should not obscure this update.

6.5 9DoF and Shape Metrics

• 3D IoU: intersection over union of predicted / ground-truth 3D boxes, jointly measuring translation, rotation, and size. Suitable for tasks such as NOCS that predict category-level dimensions.
• Chamfer Distance: bidirectional nearest-neighbor distances from predicted points to ground truth and vice versa. It evaluates shape reconstruction without guaranteeing pose accuracy.

7. Experimental analysis

7.1 Pairing Datasets and Metrics

• Instance-level tables emphasize LINEMOD / LM-O ADD(S)-0.1d and YCB-Video ADD-S AUC.
• Category-level tables commonly report CAMERA25 / REAL275 5°5cm and 3D IoU because tasks also include size / shape variation.
• BOP tables must identify the task: known-object localization often reports AR; novel-object and 6D detection protocols also require onboarding conditions and AP context.

7.2 Do Metrics Capture Application Performance?

Application question	More diagnostic metrics	What is still missing?
Instance pose under occlusion	VSD + BOP AR	Depth availability, runtime, front-end detection errors
Grasping / insertion	MSSD, translation error	Collision, gripper reachability, friction, downstream success
AR overlay / image alignment	MSPD, 2D Projection	Rendered appearance, temporal stability
Category-level pose + size	5°5cm, 3D IoU	Shape bias on unseen instances, scale definition
Online shape completion	CD + pose metrics	Surface topology, semantic parts, actual task benefit

7.3 Are Contributions Supported by Evidence?

Claimed contribution	Supporting material	My assessment
Covers three major pose formulations	Method-review sections and representative-method tables	Broad coverage, useful as a field map
Comprehensive metric / benchmark reference	Section 2.4 and task tables	Valuable metric taxonomy, but exact thresholds still require original benchmarks
Helps select appropriate methods	Cross-classification of inputs, training, inference, and applications	Helps rule out unfair comparisons; does not replace application evaluation

8. Critical analysis

Strengths

• Handles differences in problem settings and evaluation together, going beyond a collection of architecture terms.
• Unified metric notation and formulas support rapid learning and cross-linking on this site.
• Category-level and unseen-object coverage broadens a perspective otherwise limited to CAD-known instance-level tasks.

Weaknesses

• A survey cannot replace annually updated BOP task configurations and toolkit documentation.
• Statements about what is current depend on the paper version; this site should use official BOP pages for protocols and the survey for taxonomy.
• Practical measures such as speed, memory, robustness, and grasp success are not core pose errors but determine usability. Report them separately on paper pages rather than forcing them into pose accuracy.

Implicit Assumptions

• CAD or comparable shape representations are available; fully open scenes without models change the applicability of ADD / BOP.
• Benchmark object symmetries, camera intrinsics, and ground truth are trustworthy.
• Input modalities, front-end detectors / masks, and runtime measurements can be aligned across papers; this is often only partially true.

9. Implementation on This Site

• Connect papers to metric detail pages through canonical metricId, preserving original metrics strings.
• Metric detail pages need a quick summary, formulas, computation steps, strengths, interpretation boundaries, sources, and related papers.
• Use a layered catalog: 3DoF components; model points / projection; BOP errors and aggregates; 9DoF / shape.
• Explain BOP relationships in both directions: VSD / MSSD / MSPD are errors, AR aggregates localization, and AP aggregates detection.

10. Open questions

• When adding each paper, check its AUC limit, ADD(S) symmetry split, and exact data split.
• BOP-toolkit thresholds, dataset cores, and tasks may vary by year; detail pages should keep protocol versions explicit.
• For texture-defined symmetry, whether ADD-S / geometric symmetry matches actual application needs remains a dataset- and task-specific question.