Reference Pose-Rotation Pair

• Reference pose-rotation pair is a combined representation of an object's or scene's pose and rotation that formalizes spatial relationships using rigid-body geometry.
• It underpins methods in pose estimation, motion retargeting, and feature aggregation by encoding both local invariance and global pose context.
• This abstraction resolves ambiguities due to symmetry and enhances rotation-invariant architectures in tasks like 3D correspondence and dynamic kernel synthesis.

A reference pose-rotation pair is a concept central to 3D vision, robotics, and learning-based geometric perception—referring to the mathematical and algorithmic representation of an object's (or scene's, or robot's) spatial configuration, typically defined by a pose (encapsulating position and/or orientation) in a reference frame, paired with a rotation (often as a matrix, quaternion, or vectorial representation) that locates or aligns this pose relative to another pose or to a canonical frame. This abstraction underpins numerous frameworks for rotation-invariant learning, pose estimation, motion retargeting, and geometric correspondence because it enables unambiguous encoding, matching, and aggregation of local and global geometric information.

1. Mathematical Formulation of Pose–Rotation Pairs

A reference pose–rotation pair formalizes the spatial relationship between two coordinate frames, patches, points, or articulated structures via rigid-body geometry. The canonical case arises in $\mathbb{R}^3$:

• Pose: Typically denoted $\mathcal{P} = (R, t) \in \mathrm{SE}(3)$, where $R \in \mathrm{SO}(3)$ is a rotation matrix (or its equivalent) and $t \in \mathbb{R}^3$ is a translation.
• Rotation Pair: In tasks requiring relative orientation (e.g., patch-based point cloud processing, retargeting), one may consider a pair of rotation matrices $(R_1, R_2)$, each attached to a local frame (e.g., center and neighbor patch, robot and human pose).

In geometric deep learning and pose estimation, this representation is further specialized: for example, a local reference frame (LRF) at a point $p_r$ is constructed, yielding $$\mathcal{L}_r = [\partial_r^1, \partial_r^2, \partial_r^3] \in \mathbb{R}^{3 \times 3}$$. The relative pose between two such LRFs (say at $p_r$ and $p_j$) can be expressed as the 6-DoF transformation encoded by their pose–rotation pair (Chen et al., 2022).

2. Encoding Relative Geometry: Features and Representations

To encode the relative pose–rotation information in a form amenable to learning algorithms or optimization, several descriptors and parameterizations are employed:

• Augmented Point Pair Feature (APPF): For points $p_r$, $\mathcal{P} = (R, t) \in \mathrm{SE}(3)$0, and local frames $\mathcal{P} = (R, t) \in \mathrm{SE}(3)$1, $\mathcal{P} = (R, t) \in \mathrm{SE}(3)$2, the APPF $\mathcal{P} = (R, t) \in \mathrm{SE}(3)$3 encodes invariant distances, principal angles ($\mathcal{P} = (R, t) \in \mathrm{SE}(3)$4), and azimuths ($\mathcal{P} = (R, t) \in \mathrm{SE}(3)$5, $\mathcal{P} = (R, t) \in \mathrm{SE}(3)$6), forming a rotation-invariant but information-complete descriptor (Chen et al., 2022).
• Flexible Vector-Based Rotation (FVR): Represent a rotation $\mathcal{P} = (R, t) \in \mathrm{SE}(3)$7 not as a matrix, but as a pair of orthogonal vectors $\mathcal{P} = (R, t) \in \mathrm{SE}(3)$8; these are then orthonormalized (e.g., via Gram–Schmidt) to recover $\mathcal{P} = (R, t) \in \mathrm{SE}(3)$9, supporting easier regression and better learning dynamics (Chen et al., 2022).
• 6D Rotation Representations: Using two 3D vectors (columns of $R \in \mathrm{SO}(3)$0), this minimizes ambiguity and circumvents normalization constraints associated with quaternions or Euler angles; crucial for robust canonicalization and motion retargeting (Ekanayake et al., 27 Sep 2025, Figuera et al., 2024).

These representations serve as the backbone for dynamic kernel generation in convolution (PaRI-Conv), description of patchwise relations (PaRot), or decoupling local and global pose content for task performance (Chen et al., 2022, Zhang et al., 2023).

3. Role in Rotation-Invariant and Pose-Aware Architectures

Reference pose–rotation pairs resolve the fundamental tension between local rotation invariance and global pose awareness:

• Rotation-Invariance Loss: Purely invariant descriptors (e.g., based on PPFs or sorted Gram matrices) cannot distinguish globally distinct but locally symmetric structures, leading to information loss ("wing-tip feature collapse") (Guo et al., 11 Nov 2025).
• Pose-Aware Learning: By tracking or restoring reference pose–rotation pairs at each layer, architectures like PaRI-Conv and SiPF preserve full 6-DoF relative pose while maintaining rotation invariance of the output (Chen et al., 2022, Guo et al., 11 Nov 2025). Inputs are mapped to invariant descriptors (e.g., $R \in \mathrm{SO}(3)$1), which parameterize dynamic kernel synthesis, attention mechanisms, or hierarchical pooling.

For example, in PaRI-Conv (Chen et al., 2022), the dynamic convolutional kernel for a neighbor $R \in \mathrm{SO}(3)$2 is generated from the APPF: $R \in \mathrm{SO}(3)$3, where $R \in \mathrm{SO}(3)$4. The approach is provably equivariant, compact, and empirically outperforms invariant-only approaches.

4. Applications in Pose Estimation, Retargeting, and Alignment

The reference pose–rotation pair underpins a broad range of algorithmic pipelines:

• Pose Estimation from Image Pairs or Point Clouds: In absolute and relative pose estimation, utilizing reference pose–rotation pairs (e.g., via Umeyama, DRaM, Procrustes) allows closed-form or globally optimal solutions that decouple the rotation and translation subproblems (Liu et al., 2019, Hanson et al., 24 Nov 2025).
• Motion Retargeting: In data pairing between robotic and human poses, each sample is a pair $R \in \mathrm{SO}(3)$5, with $R \in \mathrm{SO}(3)$6 and $R \in \mathrm{SO}(3)$7 as sets of joint rotations (6D encoding per Zhou et al., 2019), and the correspondence is established via IK mapping, body priors, and filtering to avoid infeasible configurations (Figuera et al., 2024).
• Patchwise and Partwise Feature Aggregation: For graph-based and hierarchical models (e.g., PaRot), pose–rotation pairs enable the computation of relative pose descriptors for intra-scale and inter-scale aggregation, restoring essential geometric context otherwise lost (Zhang et al., 2023).
• Canonicalization in Human Pose: Modules like 3DPCNet infer the viewpoint-agnostic body-centric pose by directly regressing a rotation matrix (via 6D vectors) and using its transpose to transform input skeletons to the canonical frame (Ekanayake et al., 27 Sep 2025).

These approaches all leverage the capacity of pose–rotation pairs to encode and restore relative geometric information, be it for instance-level pose matching, dense correspondence, or higher-level kinematic alignment.

5. Global vs. Local Pose–Rotation Anchoring and Symmetry

A significant challenge arises in structures or scenes involving symmetry or ambiguity in reference frames:

• Symmetry Handling: Rigid object pose must be modded out by the object’s symmetry group $R \in \mathrm{SO}(3)$8; each pose is then an equivalence class $R \in \mathrm{SO}(3)$9. Efficient Euclidean embeddings for different symmetry classes enable fast search and aggregation on the pose space (Brégier et al., 2016).
• Global Reference Anchors: For global awareness, some architectures learn or optimize a consistent global rotation ("shadow" or anchor), shared across the dataset or mini-batch (e.g., via the Bingham distribution over quaternions (Guo et al., 11 Nov 2025)), which is then used to construct pose–rotation pairs that inject global reference information in an RI-consistent manner.

This dual use (resolving local orientation for relative pose while normalizing or encoding global reference) is essential for distinguishing symmetric components, transferring motion, or ensuring metric-aware learning.

6. Methodological Innovations Leveraging Pose–Rotation Pairs

Recent works extend the impact of reference pose–rotation pairs through the following methodological directions:

• Pose Decoupling and Birotation: By introducing two independent rotation matrices and optimizing their alignment to basis transformations, birotation solutions mitigate degeneracies and improve robustness in ambiguous or ill-conditioned relative pose scenarios (Zhao et al., 4 May 2025).
• Correlation-Guided Refinement: In render-and-compare or attention-guided pipelines, correspondence is established between query and reference images via geometric attention volumes, using initial reference pose–rotation pairs for iterative, learned refinement (Kim et al., 16 May 2025).
• Self-Supervised Pose Alignment: In motion canonicalization, synthetic rotations and pose-pair supervision eliminate the need for explicit calibration or ground truth, with rotation recovery anchored on reference pose–rotation pairs and regularized via composite losses that enforce consistency (Ekanayake et al., 27 Sep 2025).

These innovations exemplify the broad methodological utility of reference pose–rotation pairs for optimization, self-supervision, and structure-preserving deep network design.

7. Summary and Impact

Reference pose–rotation pairs constitute a foundational abstraction for encoding, restoring, and aggregating both local and global geometric information across 3D perception, learning, and robotics tasks. Properly constructed pose–rotation pairs ensure rigorous rotation-invariance, robust recovery of global and relative information, resilience against symmetry-induced ambiguities, and effective transference of pose across modalities and domains. Their theoretical and practical importance is substantiated by their central role in state-of-the-art point cloud classification, part segmentation, pose estimation, motion retargeting, canonicalization, and optimization frameworks across the literature (Chen et al., 2022, Ekanayake et al., 27 Sep 2025, Chen et al., 2022, Guo et al., 11 Nov 2025, Figuera et al., 2024).