Object Pose Estimation · 2026 · CVPR · CCF A

PoseGAM: Robust Unseen Object Pose Estimation via Geometry-Aware Multi-View Reasoning

PoseGAM processes a query image and multiple posed templates with a geometry-aware multiview network. Explicit point geometry and learned geometry features directly predict pose without explicitly constructing query–template correspondences, improving average AR across five BOP datasets.

Paper Project

Research area

Object Pose Estimation

Task level

Instance-level

Object setting

Unseen objects

Input modalities

RGBquery + template images

Requirements

object masksCAD / template posesmulti-view foundation modelsynthetic training data

Full paper metadata Authors, affiliations, research area, datasets, metrics, method tags, and reading status Expand

Authors: Jianqi Chen, Biao Zhang, Xiangjun Tang, Peter Wonka
Affiliations: King Abdullah University of Science and Technology
Conference / journal: Conference · CVPR · CCF A
Research area: Object Pose Estimation
Code availability: Unavailable / not yet recorded
Reading status: Overview:Available Deep Dive:Available
Task setting: Instance-level · Unseen objects
Datasets: LM-O T-LESS TUD-L IC-BIN YCB-V
Evaluation metrics: BOP AR VSD MSSD MSPD
Method tags: Novel-object Pose Estimation Template Matching & Render Retrieval Multi-object Pose Estimation Synthetic Data & Foundation Models Absolute Pose Estimation
Requirements: object masksCAD / template posesmulti-view foundation modelsynthetic training data

Reading Notes

Overview / Deep Dive

Available src/paper-notes/posegam/coarse.md

Paper overview notes

Quickly assess the paper's scope, problem setting, core method, experimental findings, and whether to read it in depth.

Paper Reading Notes (Coarse): PoseGAM: Robust Unseen Object Pose Estimation via Geometry-Aware Multi-View Reasoning

1. Basic information

Title: PoseGAM: Robust Unseen Object Pose Estimation via Geometry-Aware Multi-View Reasoning
Authors / affiliation: Jianqi Chen, Biao Zhang, Xiangjun Tang, Peter Wonka; KAUST
Year / venue: 2026 / CVPR
Paper: CVF Open Access
Project: PoseGAM
Reading mode: Overview

2. Positioning in one sentence

Query RGB and multiple posed templates support direct unseen-object pose reasoning through explicit points and learned geometry, without constructing query–template feature correspondences.

2.1 Guide to the paper’s figures and tables

Figure / table	Purpose	Overview reading tips
Figure 1	Initial / estimated CAD poses projected into queries	Check recovered spatial arrangements
Figure 2	Multimodal inputs, point-map and feature-map encoders	Geometry projected back into views
Table 1	Five BOP datasets	Generalization / stability
Tables 2 / 3	Data and geometry ablations	Architecture versus synthetic scale
Figure 4	Qualitative projections	Occlusion, appearance, multiple objects

PoseGAM qualitative results across datasets — Original cross-dataset qualitative comparison: Query Image, Mask, MegaPose, GenFlow, GigaPose, FoundPose, and Ours, from left to right. Colored object contours compare projected estimates with real images. This is not an initial-red-plane / predicted-green-plane schematic. Source: CVF Open Access. Image copyright belongs to the authors and publisher; reproduced here for paper analysis.

3. Problem setting

Query RGB plus multiple same-object templates and their poses.
Test objects are unseen; generalization spans objects, appearance, and scenes.
Output rotation / translation in the query-camera frame, evaluated with BOP AR.

4. Method overview

Encode query, templates, and camera information jointly.
Point geometry supplies explicit 3D structure.
A geometry representation network provides higher-level features.
Project geometry features to view maps compatible with foundation-image reasoning.
Predict poses end-to-end without hand-built matches / separate PnP.

5. Experimental overview

Over 190k synthetic objects with lighting, materials, appearance editing, and scene variation.
LM-O, T-LESS, TUD-L, IC-BIN, and YCB-V evaluation.
Reported mean AR improvement 5.1%, maximum dataset improvement 17.6%.

6. Quick assessment

Value: direct foundation-model geometric reasoning beyond explicit matching.
Strength: a simple interface combines multiview appearance and geometry.
Risks: posed templates / CAD and large-scale training remain necessary; less training-free than ConceptPose, with additional multiview preparation.
Audience: unseen pose, multiview Transformers, geometry tokens, synthetic transfer.

Available src/paper-notes/posegam/deep.md

In-depth paper notes

Support reproduction, comparison, related-work writing, and identifying opportunities for further research.

Paper Reading Notes (Deep): PoseGAM: Robust Unseen Object Pose Estimation via Geometry-Aware Multi-View Reasoning

1. Key assessment

Paper: CVF Open Access
Project: PoseGAM
Key assessment: explicit object geometry and reference / query appearance mismatch are the challenge, beyond attaching a pose head to a multiview model.

2. Positioning in one sentence

Point maps and learned geometry maps condition direct query-plus-template reasoning without an explicit correspondence stage.

2.1 Guide to the paper’s figures and tables

Figure / table	Deep-dive questions
Figure 1	Do CAD projections recover both position and orientation?
Figure 2	How are geometry / features encoded and projected into views?
Table 1	Are five-dataset gains concentrated in certain materials / scenes?
Figure 4	Do masks or reasoning cause appearance / occlusion / clutter failures?
Ablations	Are explicit and learned geometry complementary?

PoseGAM qualitative comparisons across datasets — Original cross-dataset comparison: Query Image, Mask, and colored object projections from MegaPose, GenFlow, GigaPose, FoundPose, and Ours. It compares methods rather than color-coded initial / predicted pose planes. Source: project / CVF Open Access. Image copyright belongs to the authors and publisher; reproduced here for paper analysis.

3. Problem setting and task boundaries

Inputs: query RGB, template RGBs, and known template poses / CAD geometry.
Outputs: target 6D poses in the query frame.
Unseen test objects require learned geometry / reasoning rather than instance memorization.
Queries may contain multiple objects; Figure 1 emphasizes spatial arrangement.
Dependencies: masks, posed templates, and geometry make this generalizable / model-based rather than training-free / CAD-free.

4. Research context

Match-then-localize / refine expose correspondences whose errors propagate into solvers.
Multiview foundations infer cameras / geometry, but RGB-only inputs omit CAD and commonly assume consistent appearance.
PoseGAM treats geometry as a peer input, aligning it through view maps.

5. Research motivation

Removing explicit matching reduces dependence on engineered correspondences.
Raw geometry tokens disrupt image interfaces, motivating feature extraction and projection.
Synthetic data must cover CAD-to-real color, material, lighting, and background shifts.

6. Method details

6.1 Multimodal inputs

Image encoders process queries / templates; camera encoders process poses.
Point-map geometry is encoded into view-compatible features.
Learned geometry provides more stable object-level features.

6.2 Geometry-to-view map

Raw geometry-token sequences were found to hinder learning.
Extract features first, then project them into view positions for joint attention.
This retains multiview inductive structure while supplying explicit CAD evidence.

6.3 End-to-end prediction

Joint multimodal representations directly regress query poses.
Separate matching, RANSAC, PnP, and secondary refinement are omitted, reducing interfaces.
BOP metrics account for symmetric equivalence.

7. Data and training

Over 190k synthetic objects across lighting, textures, materials, and environments.
Blender standardizes asset materials and renders centric / uncentric views.
DDIM inversion / FLUX edit color, lighting, and appearance while preserving structure.
Training targets query pose; smaller subsets reduce ablation costs.
Object splits and appearance shifts matter because generalization is the goal, beyond benchmark-specific retraining.

8. Experimental analysis

8.1 Evaluation

LM-O, T-LESS, TUD-L, IC-BIN, YCB-V.
BOP AR combines VSD / MSSD / MSPD recall; ablation AUC combines absolute rotation / translation accuracy.

8.2 Main results

Table 1 mean AR is 41.1: LM-O 43.0, T-LESS 34.1, TUD-L 56.8, IC-BIN 24.3, YCB-V 47.4.
Reported mean improvement is 5.1%, maximum 17.6% on one dataset.
Multiview reasoning helps shape, clutter, and occlusion, while IC-BIN remains difficult.

8.3 Priority ablations

Remove explicit geometry to test CAD use beyond RGB.
Remove learned geometry to test structure beyond raw tokens.
Remove edited appearance to quantify transfer gaps.
Remove camera / pose embeddings to test spatial reasoning.

9. Critical analysis

Strength: clear geometry injection through reusable view-map interfaces.
Limit: multiple posed templates / reference preparation and large training data.
Implicit matching cost: attention is harder to diagnose than explicit incorrect PnP point pairs.
Scope: five BOP datasets do not establish CAD-free, template-free, or fast-video performance.

10. Implications for my research

Point maps, NOCS, Gaussians, and canonical features can be view-aligned conditions beyond token concatenation.
Combine explicit and learned geometry branches in multiview systems.
Add correspondence / uncertainty probes for diagnosis and active views.

11. Further analysis: multiview geometry reasoning

11.1 Reading the main table

Table 1 BOP AR:

Dataset	LM-O	T-LESS	TUD-L	IC-BIN	YCB-V	Mean
PoseGAM	43.0	34.1	56.8	24.3	47.4	41.1

The abstract reports mean improvement 5.1% and maximum 17.6%. Means weight five datasets equally, unlike single-dataset ADD-AUC. LM-O / YCB-V emphasize clutter / occlusion; TUD-L geometric structure; low IC-BIN indicates persistent heavy occlusion / industrial-appearance challenges.

11.2 Geometry inside the network

Geometry is repeatedly read by multiview reasoning rather than matched by nearest neighbors before PnP:

CAD / template points and camera poses form geometry.
The representation network learns structural features.
View-map projections interact with RGB / foundation features.
Camera relations aggregate template evidence for pose output.

Multiple viewpoints structurally constrain one pose, trading explicit correspondence diagnostics for implicit reasoning.

11.3 Data and ablation priorities

Over 190k objects with lighting, materials, environments, and FLUX / DDIM appearance editing.
First remove point geometry to test CAD structure use.
Next remove learned geometry to isolate high-level encoding.
Then reduce templates / corrupt poses to test view reasoning.
Finally reduce edited appearances; large drops would reveal dependence on synthetic-to-real coverage beyond geometry.

11.4 Claims and analysis

Level	Assessment
Authors’ claim	Geometry-aware reasoning improves five unseen-object BOP datasets.
Evidence	Mean AR 41.1, maximum gain 17.6%, and multi-object projections.
Reservation	Posed templates are strong priors, unlike single-reference / image ConceptPose and OrienPose settings.
Next tests	Template count, pose noise, masks, incomplete CAD, shuffled views, geometry attention, and residuals.

12. Reproduction checklist

Specify query, template, CAD, and camera frames, including camera-to-object versus object-to-camera.
Fix BOP thresholds / symmetries for reproducible AR.
Record training and test template counts separately; fewer test views may shift the input distribution.
Stratify low texture, repetition, over-half occlusion, and mask errors beyond five-dataset means.

Related papers

2026 · CVPR

3PT

3PT unifies detection, segmentation, and 6-DoF pose with two Transformers for 2D / 3D object-centric understanding. Continuous refinement of object representations and rendered-feature alignment support zero-shot cross-object perception on BOP-H3, BOP-Industrial, and BOP-Classic.

2026 · AAAI

6DAttack

Studies backdoor attacks on PnP-based and end-to-end 6-DoF pose pipelines. Structurally 3D triggers and modified pose labels are injected into training data so that models behave normally on clean inputs but produce controlled incorrect poses when triggered.

2026 · ICRA

Active Tactile Exploration

A violation-implicit loss incorporates physical-constraint violations into joint estimation of freely moving unknown convex rigid-body trajectories and cuboid / 20-vertex polytope shapes. Observed / Fisher information at a local MLE estimates EIG to select the next touch. In simulation, the EIG strategy yields stable errors below 4 mm after 5 cuboid actions; real Trifinger results at step 5 substantially outperform random selection. Convex geometry, local optima, multiple initializations, and tactile / proprioceptive errors limit the method.