A Geometry-Driven, Framework-Agnostic Optimization for Object Pose Estimation

Wei Chen; Tao Zhen; Zhongchen Shi; Jing Zhang; Liang Xie; Erwei Yin

cs.CVAug 27, 2026

A Geometry-Driven, Framework-Agnostic Optimization for Object Pose Estimation

Authors: Wei Chen, Tao Zhen, Zhongchen Shi, Jing Zhang, Liang Xie, Erwei Yin

Abstract

Current object pose estimation research remains predominantly model-centric, focusing on architectural innovations and post-processing refinements. This paper introduces a data-centric optimization by proposing a novel, physically grounded rotation representation through principal axes alignment. Our method aligns the object's coordinate system with its inherent geometric axes, derived from inertial properties, yielding three key advantages: Inherent Stability-leveraging the energy-minimizing property of principal axes provides a robust representation that is less sensitive to noise and occlusions; Symmetry-Aware Canonicalization-explicitly resolving rotational ambiguities for symmetric objects at the data level, which fundamentally eliminates label confusion during network training; and Framework Agnosticism-the optimization is applied purely at the dataset level, ensuring plug-and-play compatibility with existing networks without any architectural modification. We validate the framework across diverse category-level and instance-level models. Extensive experiments demonstrate consistent and significant accuracy improvements, while preserving the integrity of the baseline network. This work establishes a new, geometry-driven direction for enhancing pose estimation, circumventing the need for complex network redesign.

Figures & tables

Fig. 1: Semantic illustration of proposed mechanism .The main contribution of previous methods mainly focused on network layout update, including new loss functions or new network architectures. Our proposed method only updates the rotation coordinate for pose representation, which is easy to apply and perfectly compatible with existing methods.

Fig. 2: Overview of the proposed mechanism and its usage. There are four main steps of our proposed mechanism. First, we calculate the principal axes of the object’s inertia, and then we form the transformation rotation using these three principal axes. Third, we access the new object coordinate for the datasets with the transformation rotation. Then, we eliminate the rotation ambiguity with the proposed module if the objects are symmetric. Finally, we retrain the network using datasets in new coordinates, eliminating rotation ambiguity.

Fig. 3: Physics property of $\displaystyle\pi$ -Mechanism . For original coordinates, the non-diagonal elements are non-zero. For the object in Principal Axes, the non-diagonal elements are close to zero.

Fig. 4: Illustration of the effect of

\displaystyle\pi

-Mechanism . The above row shows the geometric relation between the object and its coordinates. When the rotation value is 0, the orientation of the object is not aligned well with the axes of the object coordinate. The second row of images shows the relation when applying our proposed mechanism to the original object relation. The orientation of the object is aligned well with its coordinate.

Fig. 5: A Taxonomy of 3D Objects Based on Rotational Symmetry Around Two Principal Axes . We systematically categorize objects’ shapes by their rotational symmetry properties around two independent principal inertial axes. The rows and columns represent the four fundamental symmetry types for a single axis: Continuous, N-fold Discrete, 2-fold, and Asymmetric. Each cell provides one or two canonical 3D object examples (without considering the color texture) exhibiting the corresponding pairwise combination of symmetries.

Fig. 6: Pipeline for resolving rotational ambiguity in symmetric objects. Left top: Computation of the principal axes of inertia from the object’s geometry. Right top: Symmetry type classification (e.g., two-fold symmetry) based on the principal axes. Right bottom: Generation of canonical candidate vectors. Left bottom: Final unambiguous rotation representation obtained after disambiguation.

Fig. 7: Illustration of rotation disambiguation for symmetric objects. Left column: Distribution of rotation viewpoints in the original ambiguous space. Middle column: 3D model of the object in the TLESS dataset (obj29, obj27, obj14), highlighting its intrinsic symmetry. Right column: Unique distribution of rotation viewpoints after ambiguity resolution.

Fig. 8: Illustration of Geometric Vectors for Different Symmetry Types.

Method	Ape	Bench Vise	Camera	Can	Cat	Driller	Glue	Hole Puncher	Iron	Lamp	Phone	Average
PointNet [ 43 ]	12.28	12.49	11.76	13.71	13.11	11.30	10.72	10.23	11.62	10.47	11.09	11.71
PointNet + $\displaystyle\pi$ -Mechanism	12.12	10.30	10.27	13.16	12.51	11.26	11.41	12.34	11.42	10.01	9.71	11.32
Si-Mamba [ 44 ]	10.85	11.04	10.42	11.34	11.22	11.06	10.13	12.55	10.17	11.57	11.34	11.06
Si-Mamba + $\displaystyle\pi$ -Mechanism	10.48	10.58	10.05	11.94	10.26	10.49	9.86	11.46	10.67	10.25	10.72	10.61

TABLE I: Rotation error for different methods on the synthetic LINEMOD dataset. The best result is bolded.

Method	Ape	Bench Vise	Camera	Can	Cat	Driller	Duck	Hole Puncher	Iron	Lamp	Phone	Average
G2L [ 50 ]	7.37	15.34	5.99	8.96	6.03	13.04	6.84	5.13	17.20	12.54	7.24	9.61
G2L + $\displaystyle\pi$ -Mechanism	6.25	11.94	5.83	6.30	5.23	9.37	6.58	3.94	7.37	9.79	5.44	7.09
GDR-Net [ 27 ]	2.11	1.85	1.81	1.82	2.02	2.02	1.98	1.96	2.33	1.87	2.30	2.01
GDR-Net+ $\displaystyle\pi$ -Mechanism	1.83	1.82	1.67	1.73	1.80	1.83	1.83	1.88	2.04	1.68	2.19	1.83

TABLE II: Rotation error for different methods on LINMOD dataset. The best result is bolded.

Fig. 9: Mean rotation errors with standard deviation (SD). The left bar is the mean rotation error and SD of the baseline method; The right bar is the mean rotation error and SD of the proposed mechanism.

Panel of Figure 12 — Fig. 10: The Quantitative results for the LINEMOD dataset. The green boxes are the ground truth. The red boxes are our results. The blue boxes are the results of the baseline.

Methods	Ape	Can	Cat	Driller	Duck	Egg box	Glue	Hole Puncher	Average
LC	44.44%	89.06%	49.87%	87.81%	56.08%	62.81%	68.88%	72.89%	66.69%
LC+ $\displaystyle\pi$ -M	48.97%	90.22%	51.39%	89.46%	57.31%	65.37%	70.86%	76.61%	68.99%

TABLE III: Results on dataset LINEMOD-OCCLUDE for method GDR-LC [ 14 ] in ADD(-S)-0.1d metric (the higher is better). The best result is bolded. For symmetric objects ‘Egg Box’ and ‘Glue’, the ADD-S scores are reported.

Panel of Figure 14 — Fig. 11: The Quantitative results for LINEMOD-OCC dataset. The green boxes are the ground truth. The red boxes are our results. The blue boxes are the results of the baseline.

Method	obj 14	obj 27	obj 29
G2L [ 50 ]	31.51	71.70	108.83
G2L + $\displaystyle\pi$ -Mechanism	2.64	10.11	17.79

TABLE IV: Rotation error for different methods on TLESS symmetric objects. The best result is bolded.

Fig. 12: ‘Camera’ category with different instances

	Category	Bottle	Bowl	Can	Camera	Laptop	Mug	Average
different representation	FS-Net	5.71	3.85	3.97	21.59	8.00	13.77	9.48
different representation	FS-Net+ $\displaystyle\pi$ -Mechanism	5.34	3.64	3.58	20.99	7.57	12.40	8.92

TABLE V: Rotation error comparison. Results on dataset NOCS-REAL for method FS-Net [ 6 ] in rotation prediction error. The best result is bolded.

Fig. 13: Loss value comparison . Left: the loss value comparison between original datasets and datasets in Principal Axes coordinates. Right: the loss ratio between original and principal over training iterations.

Symmetry-Aware Object Pose Estimation

See all 11 papers

UniPose9D: Universal Category-Agnostic Object Pose Estimation
Jul 10, 2026Yang You, Yi Du, Cole Harrison +1Object Pose Estimation
Stanford University · Amazon
PriorPose: Reference-Guided Joint Deformation and Alignment for Category-Level Object Pose Estimation
Sep 15, 2026Yihan Chen, Huan Ren, Wenfei Yang +3Object Pose Estimation
School of Information Science and Technology / National Key Laboratory of Deep Space Exploration, University of Science and Technology of China · Beijing Institute of Control Engineering
Symmetry-Aware 9D Pose Estimation with Sim(3)-Consistent Feature and Spherical Inception Convolution
Jun 1, 2026Panfei Cheng, Hongshan Yu, Wenrui Chen +3Object Pose Estimation Convolutional Neural Networks
National Engineering Research Center for Robot Visual Perception and Control, School of Robotics and Artificial Intelligence, Hunan University, Changsha 410012, China

Jul 10, 2026J/K move · Enter open · S save

UniPose9D: Universal Category-Agnostic Object Pose Estimation

Yang You, Yi Du, Cole Harrison, Leonidas Guibas

Stanford University · Amazon

Object Pose Estimation

Open paper AskPDF arXiv

Object pose estimation is a fundamental problem in 3D vision. Although recent state-of-the-art approaches achieve strong performance, generalization to novel categories and unseen scenes remains challenging. We propose UniPose9D, a unified model for category-agnostic 9D object pose estimation: given an instance mask/ROI and either an RGB-D observation or an RGB image with predicted depth, the model estimates rotation, translation, and metric size without category labels, CAD models, mean-shape priors, or reference views. Specifically, UniPose9D samples point pairs from the observed object geometry and uses DINOv2 and PointNet features to predict NOCS coordinates for each pair. To improve accuracy, we introduce a point-pair-based RANSAC N-hop Kabsch-Umeyama algorithm with an adaptive threshold. We further employ flow matching to address symmetric ambiguities and construct a large-scale training set by curating and aligning pose annotations from existing public datasets. Experiments across eight datasets show that a single unified model achieves competitive performance on standard benchmarks while generalizing to unseen objects, unseen categories, and in-the-wild scenarios. Our code and model are available at https://github.com/qq456cvb/UniPose9D.

Abstract

Symmetry-Aware Object Pose Estimation

Explore similar work

UniPose9D: Universal Category-Agnostic Object Pose Estimation

PriorPose: Reference-Guided Joint Deformation and Alignment for Category-Level Object Pose Estimation

Symmetry-Aware 9D Pose Estimation with Sim(3)-Consistent Feature and Spherical Inception Convolution