Enhancing Rotation-Invariant 3D Learning with Global Pose Awareness and Attention Mechanisms

TL;DR

This work addresses the limitation of rotation-invariant 3D learning where global pose cues are lost due to restricted receptive fields, causing indistinguishability of symmetric components (wing-tip collapse). It introduces Shadow-informed Pose Feature (SiPF), which augments local RI descriptors with a globally consistent shadow derived from a learnable rotation, and couples it with Rotation-invariant Attention Convolution (RIAttnConv) and a Bingham-distribution-based shadow locating module to preserve global pose information while maintaining RI. The approach yields state-of-the-art results on 3D classification and part segmentation under arbitrary rotations, demonstrating strong gains especially in fine-grained spatial discrimination and robustness to real-world noise. This method enhances practical 3D perception by enabling rotation-invariant processing that remains sensitive to global structure, with potential impact on autonomous systems and robotics where objects appear in unconstrained orientations.

Abstract

Recent advances in rotation-invariant (RI) learning for 3D point clouds typically replace raw coordinates with handcrafted RI features to ensure robustness under arbitrary rotations. However, these approaches often suffer from the loss of global pose information, making them incapable of distinguishing geometrically similar but spatially distinct structures. We identify that this limitation stems from the restricted receptive field in existing RI methods, leading to Wing-tip feature collapse, a failure to differentiate symmetric components (e.g., left and right airplane wings) due to indistinguishable local geometries. To overcome this challenge, we introduce the Shadow-informed Pose Feature (SiPF), which augments local RI descriptors with a globally consistent reference point (referred to as the 'shadow') derived from a learned shared rotation. This mechanism enables the model to preserve global pose awareness while maintaining rotation invariance. We further propose Rotation-invariant Attention Convolution (RIAttnConv), an attention-based operator that integrates SiPFs into the feature aggregation process, thereby enhancing the model's capacity to distinguish structurally similar components. Additionally, we design a task-adaptive shadow locating module based on the Bingham distribution over unit quaternions, which dynamically learns the optimal global rotation for constructing consistent shadows. Extensive experiments on 3D classification and part segmentation benchmarks demonstrate that our approach substantially outperforms existing RI methods, particularly in tasks requiring fine-grained spatial discrimination under arbitrary rotations.

Figures (6)

• Figure 1: Illustration of pose ambiguity caused by symmetric local structures. Traditional Point Pair Features (PPF) capture only local geometric relationships and thus yield identical representations for symmetric regions with similar local geometry (e.g., left and right wing tips). In contrast, the proposed SiPF incorporates global pose awareness by introducing a 'shadow' reference point derived from a shared rotation matrix $R_g$ to incorporate pose-aware context, enabling consistent and discriminative RI representations across arbitrary orientations.
• Figure 2: Overview of the proposed SiPF and RIAttnConv pipeline. Our framework consists of three main components: (1) Task-adaptive Shadow Locating estimates a globally consistent reference rotation $R_g$ by modeling unit quaternions with a Bingham distribution, generating 'shadow' points for each input; (2) SiPF and Input Feature Extraction constructs LRFs $\{\mathcal{L}_j\}^k_{j=1}$ and computes SiPFs $\mathbf{P}_r$ by combining local geometric descriptors with global shadow-based pose differences; (3) RIAttnConv embeds SiPFs to produce adaptive kernel weights $\mathbf{W}_r$, which are used in a RI attention block to guide feature aggregation. A reversed EdgeConv module then fuses neighbourhood and center features $x_r$ to produce the final output $x_r'$.
• Figure 3: Visualization of part segmentation results on ShapeNetPart Dataset under z/SO(3) setting. The leftmost column is the ground truth. The testing results under arbitrary rotations are on the rest columns.
• Figure 4: Illustration of the Degeneracy under Shadow–Axis Alignment. When the shadow point $p_r'$ lies along the primary LRF axis $\partial_1(p_r)$, the pairwise geometric descriptor $\textnormal{PPF}(p_r, p_r')$ becomes constant, causing SiPPF to degenerate into standard PPF, which is known to be ambiguous to azimuthal variations around $\partial_1(p_r)$. (a) Reference point $p_r$ and its shadow $p_r'$ are vertically aligned, and neighbours $p_j$ and $p_j'$ lie on the circular ring centered at $p_r$. (b) A different neighbour $p_j$ with rotated azimuth but the same distance and normal still yields the same $\textnormal{PPF}(p_r,p_j)$. (c) Due to the constancy of $\textnormal{PPF}(p_r,p_r')$, the SiPPF representation collapses to $\textnormal{PPF}(p_r,p_j)$, thereby losing discriminability among neighbours with different angular positions around $\partial_1(p_r)$.
• Figure 5: Illustration of the Degeneracy under Shadow–Local Coincidence. When the global rotation $R_g$ coincides with the patch-swapping transformation $R_j$, the SiPPF descriptor becomes insensitive to residual in-plane rotations of $p_j$, failing to distinguish the local receptive field $\Omega(p_j)$ from its rotated counterpart $\Omega(p_j) R_j$. (a) The local reference point $p_r$ is rotated to $p_r^R$ through the local transformation $R_j$, aligning the LRF with the receptive field. (b) The global rotation $R_g$ is equal to $R_j$, resulting in the shadow $p_r' = p_r^R$. (c) The inverse of $R_g$ is applied to $p_r^R$, resulting in $p_r^{'R} = p_r$. This makes the SiPF invariant to $R_j$, hence losing its ability to distinguish between the original receptive field $\Omega(p_j)$ and the rotated variant $\Omega(p_j) R_j$.

Theorems & Definitions (16)

• Lemma 1
• proof
• Theorem 1
• proof
• Theorem 2
• proof
• Lemma 2
• proof
• Theorem 3
• proof