COG: Confidence-aware Optimal Geometric Correspondence for Unsupervised Single-reference Novel Object Pose Estimation

Yuchen Che; Jingtu Wu; Hao Zheng; Asako Kanezaki

cs.CVFeb 28, 2026

COG: Confidence-aware Optimal Geometric Correspondence for Unsupervised Single-reference Novel Object Pose Estimation

Authors: Yuchen Che, Jingtu Wu, Hao Zheng, Asako Kanezaki

Organizations: Institute of Science Tokyo · RIKEN · Tohoku University

Abstract

Estimating the 6DoF pose of a novel object with a single reference view is challenging due to occlusions, view-point changes, and outliers. A core difficulty lies in finding robust cross-view correspondences, as existing methods often rely on discrete one-to-one matching that is non-differentiable and tends to collapse onto sparse key-points. We propose Confidence-aware Optimal Geometric Correspondence (COG), an unsupervised framework that formulates correspondence estimation as a confidence-aware optimal transport problem. COG produces balanced soft correspondences by predicting point-wise confidences and injecting them as optimal transport marginals, suppressing non-overlapping regions. Semantic priors from vision foundation models further regularize the correspondences, leading to stable pose estimation. This design integrates confidence into the correspondence finding and pose estimation pipeline, enabling unsupervised learning. Experiments show unsupervised COG achieves comparable performance to supervised methods, and supervised COG outperforms them.

Figures & tables

Figure 2: Pre-processing pipeline of COG. Given RGB-D inputs, an object is segmented, depth map is back-projected into point clouds, and per-point RGB features are extracted from DINO to form feature augmented inputs.

Figure 3: Overview of the COG framework. The pipeline consists of coarse and fine phases, each using a geometric transformer to predict point-wise confidences and features. A Sinkhorn-based OT module computes soft correspondences, and a weighted SVD solver estimates the rigid transformation. The coarse pose is further refined in the fine phase using position embeddings for precise alignment.

Method	Supervision	Modality	Reference	LM-O [ 3 ] $\uparrow$	TUD-L [ 29 ] $\uparrow$	YCB-V [ 70 ] $\uparrow$	Mean $\uparrow$	Time (s) $\downarrow$
PPF [ 16 ]	None	PC	Image	29.7	14.8	38.3	27.6	11.8
FPFH+MAC [ 54 , 75 ]	None	PC	Image	22.5	22.1	49.6	31.4	136.9
FPFH+RANSAC [ 54 , 19 ]	None	PC	Image	31.0	31.0	50.0	37.3	6.4
PPF+ICP [ 16 , 2 ]	None	PC	Image	44.7	29.1	66.8	46.9	14.3
Robust OT [ 55 ]	None	PC+RGB	Image	45.5	66.3	66.0	59.3	4.0
FreeZe [ 4 ]	None	PC+RGB	Image	45.5	68.3	65.5	59.8	53.0

Table 1: Quantitative comparison of single-reference novel object pose estimation methods on LM-O, TUD-L, and YCB-V. COG achieves state-of-the-art performance in both supervised and unsupervised settings, with comparable inference speed.

Figure 4: Qualitative results of unsupervised COG on LM-O, TUD-L, and YCB-V datasets. Blue bounding boxes represent the estimated poses, while white boxes denote ground-truth poses.

Figure 5: Visualization of predicted confidence from unsupervised COG. Our method effectively handles non-overlapping regions and outlier points by assigning low confidence to unreliable points. Poses are aligned for visualization.

Method	Dragon	Frog	Watering Can	Mean
UnoPose [ 41 ]	70.0	72.2	59.1	67.1
COG (Supervised)	72.9	68.3	83.9	75.0
COG (Unsupervised)	71.2	64.4	81.2	72.3

Table 2: Overlapping IoU results on TUD-L benchmark’s objects.

	Correspondence	$\mathcal{L}_{\mathrm{sem}}$	$\mathcal{L}_{\mathrm{cycl}}$	VSD $\uparrow$	MSSD $\uparrow$	MSPD $\uparrow$	Mean $\uparrow$
A1)	Argmax			71.3	78.4	64.4	71.4
A2)	Argmax		✓	72.1	78.5	64.2	71.6
A3)	Argmax	✓		68.3	79.1	64.0	70.5
A4)	Argmax	✓	✓	73.4	80.0	65.9	73.1
S1)	Softmax			71.9	78.6	64.0	71.5
S2)	Softmax		✓	72.1	78.5	64.1	71.6

Table 3: Modules and losses ablation on YCB-V benchmark. Correspondence refers to whether use argmax, softmax, uniform marginal OT, or confidence marginal OT to estimate correspondence,

\mathcal{L}_{\mathrm{sem}}

and

\mathcal{L}_{\mathrm{cycl}}

refer to the use of these loss terms.

	Sinkhorn Iteration	Semantic Priors	mAP $\uparrow$	ENT $\downarrow$	$\Delta\mathrm{Marginal}$ $\downarrow$
C4-1g)	1		73.1	21.1	0.35
C4-2g)	2		73.2	23.0	0.27
C4-4g)	4		73.2	24.1	0.21
C4-8g)	8		73.0	25.7	0.16
C4-1s)	1	✓	75.8	9.2	0.48
C4-2s)	2	✓	75.9	10.5	0.38

Table 4: OT parameters ablation on YCB-V benchmark. Effective Number of Tokens (ENT) reflects the number of points receiving attention, lower values indicate sharper focus.

\Delta\mathrm{Marginal}

indicates the distance between target marginals and result marginals.

Figure 6: Performance vs runtime (including segmentation, DINO feature extraction, and pose estimation). $t$ indicates the number of refinement iterations.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

\mathbf{\Pi}_{[i,j]}=\exp\!\left(\log\mathbf{K}_{[i,j]}+\mathbf{u}_{[i]}+\mathbf{v}_{[j]}\right).

Appendix

Algorithm 1 Log-Domain Sinkhorn with Target Marginals

Figure 7: t-SNE results of DINO features on query and reference $\mathbf{F}_{p},\mathbf{F}_{q}$ , and denoised semantic features $\mathbf{S}_{p},\mathbf{S}_{q}$ , colored by k-means part segmentation. With semantic denoising, the distance between points from different views (triangle and circle) with the same semantic part is reducing dramatically.

Figure 8: Visualization of pseudo confidence label generation. Points with smaller Chamfer, cycle, and semantic distances receive higher confidence labels.

Method	LM-O [ 3 ]	TUD-L [ 29 ]	YCB-V [ 70 ]	Mean $\uparrow$
DINO+PHS [ 48 ]	24.7	16.2	45.2	28.7
COG (1% data)	52.9	66.9	73.6	64.5
COG (5% data)	55.7	70.5	75.7	67.3
COG (25% data)	55.1	70.7	75.4	67.2
COG (50% data)	55.4	71.8	74.7	67.3
COG (100% data)	56.7	73.8	75.9	68.8

Appendix

Table 5: Quantitative comparison of DINO with pose-hypothesis-scoring (PHS) and COG trained with different fractions of the training data.

Figure 9: Relationship between predicted confidence and point-level distances for all test query-reference pairs in TUD-L [ 29 ] benchmark. The geometric distance (log scale) and semantic cosine distance are measured between each point and its nearest neighbor.

Figure 10: Per-object mean performance and overlapping.

Benchmark	Object	VSD $\uparrow$	MSSD $\uparrow$	MSPD $\uparrow$	Mean $\uparrow$
LM-O [ 3 ]	ape	40.0	54.7	65.5	53.4
	can	51.1	70.2	68.0	63.1
	cat	37.6	53.1	58.1	49.6
	driller	47.7	56.1	51.6	51.8
	duck	53.1	59.2	68.6	60.3
	eggbox	42.9	59.8	63.6	55.4

Appendix

Table 6: Per-object results on the LM-O [ 3 ] , TUD-L [ 29 ] , and YCB-V [ 70 ] benchmarks.

Figure 11: Failure cases of our method. (a) indicates the segmentation failure where segmentation mask is incorrect, and (b) indicates the unsupervised limitation discussed in Sec. C.2 .

Figure 12: Results on 3 BOP benchmarks. Blue bounding boxes represent the estimated poses, while white boxes denote ground-truth poses.

Object Pose Estimation

Object pose estimation recovers an object's 3D position and orientation from images, depth, point clouds, or contact sensing.

See all 187 papers

Learning Cross-View Semantic Priors for Single-Reference Unseen Object Pose Estimation
Jun 20, 2026Jiahong Chen, Jinghao Wang, Ziwen Wang +3Multi-View Consistency Object Pose Estimation
College of Aerospace Science and Engineering, National University of Defense Technology, Changsha 410073, China · Hunan Provincial Key Laboratory of Image Measurement and Vision Navigation, Changsha 410073, China
Object Pose Transformer: Unifying Unseen Object Pose Estimation
Mar 24, 2026Weihang Li, Lorenzo Garattoni, Fabien Despinoy +2Object Pose Estimation Relative Pose Estimation
Technical University of Munich · Munich Center for Machine Learning · Toyota Motor Europe
Pose Anything Anywhere:Model-free Object Poses from Arbitrary References
Jun 22, 2026Hongli Xu, Jiaqi Hu, Junwen Huang +5Multi-View Geometry Object Pose Estimation
Technical University of Munich · Siemens AG · Munich Center for Machine Learning (MCML) +2

Jun 20, 2026J/K move · Enter open · S save

Learning Cross-View Semantic Priors for Single-Reference Unseen Object Pose Estimation

Jiahong Chen, Jinghao Wang, Ziwen Wang, Zi Wang+2

College of Aerospace Science and Engineering, National University of Defense Technology, Changsha 410073, China · Hunan Provincial Key Laboratory of Image Measurement and Vision Navigation, Changsha 410073, China

Multi-View Consistency Object Pose Estimation

Open paper AskPDF arXiv

Single-reference unseen object 6D pose estimation reduces object onboarding by estimating poses of arbitrary novel objects from only one reference view. Recent correspondence-based pipelines have achieved robust performance with vision foundation model (VFM) features. However, they typically treat these features as intra-view descriptors, leaving dense visual-semantic cues, including appearance, structure, and context, insufficiently exchanged across views before geometric decoding. Consequently, the decoded point features may lack joint semantic and geometric discriminability, making correspondence estimation still difficult in challenging cases. Instead of processing features independently, we build the correspondence pipeline around an early cross-view semantic prior. Specifically, cross-view semantic interaction (CVSI) enables dense query and reference VFM tokens to exchange semantic context and form a cross-view prior. Nevertheless, direct CVSI may disturb the VFM token structure, while the resulting semantic prior still needs 3D representation consistency for rigid correspondence. To make this CVSI prior reliable for 3D correspondence learning, we introduce two complementary training-time constraints: the intra-view structure preservation (IVSP) loss preserves the original intra-view token affinity structure during interaction, while the reference-anchored geometric consistency (RAGC) loss enforces spatial representation consistency of decoded point features. The final pose is recovered from learned correspondences through weighted SVD. We further construct a challenging view-pair protocol from the BOP Challenge datasets YCB-V and TUD-L to evaluate robustness in difficult matching scenarios. Extensive experiments on six benchmarks under different view-pair settings show that our method achieves state-of-the-art performance while maintaining comparable inference speed.

Abstract

Object Pose Estimation

Explore similar work

Learning Cross-View Semantic Priors for Single-Reference Unseen Object Pose Estimation

Object Pose Transformer: Unifying Unseen Object Pose Estimation

Pose Anything Anywhere:Model-free Object Poses from Arbitrary References