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.
Introduction
In real-world applications such as robotics and autonomous driving, 3D objects are frequently encountered at arbitrary orientations. Standard deep learning models for point clouds often rely on pre-aligned or "upright" data, failing when faced with rotation. To address this, Rotation-Invariant (RI) methods have emerged, replacing raw coordinates with geometric descriptors like distances and angles.
However, these local RI methods possess a significant "blind spot" due to their restricted receptive fields. As illustrated in Figure 1, this leads to Wing-tip feature collapse: local descriptors at the tips of an airplane's left and right wings are identical. Without a global reference, the model cannot distinguish between these spatially distinct but geometrically similar structures.
Methodology
The proposed framework re-introduces global context into RI learning through two primary innovations: the Shadow-informed Pose Feature (SiPF) and the Rotation-invariant Attention Convolution (RIAttnConv).
Shadow-informed Pose Feature (SiPF)
The model generates a "shadow" point ($p'_r$) for every point in the cloud, acting as a consistent global anchor. By calculating the relationship between a point, its neighbors, and this global shadow, the model gains "global pose awareness." Since the shadow rotates with the object (via a learned global rotation $R_g$), the relative distance to the shadow remains constant, satisfying rotation invariance while providing a global "north star."
Task-Adaptive Shadow Locating
To determine the optimal shadow orientation, the authors utilize a Bingham distribution over unit quaternions. This allows the model to dynamically learn an optimal global rotation during training, ensuring shadows are placed to best support the specific task (e.g., classification vs. segmentation).
Rotation-invariant Attention Convolution (RIAttnConv)
The RIAttnConv operator integrates SiPFs into an attention mechanism. Instead of simple aggregation, it uses global-local relationships to weight neighbor features, allowing the model to focus on the most discriminative parts of the shape. This is complemented by a Reversed EdgeConv strategy that explicitly combines aggregated neighbor features with the original center point features to preserve reference point identity.
Experimental Results
The framework was evaluated on three major benchmarks under arbitrary rotation protocols:
- ModelNet40: The model achieved a state-of-the-art 92.6% accuracy under arbitrary rotations, outperforming recent RI methods such as RI-GCN (91.0%) and PaRot (90.8%).
- ShapeNetPart: In part segmentation, the model reached a class mIoU of 82.9%, effectively demonstrating its ability to resolve the wing-tip collapse problem by distinguishing symmetric parts.
- ScanObjectNN: On this challenging real-world dataset with noise and occlusions, the model reached 84.0%, significantly leading over previous benchmarks like PaRI-Conv (83.3%).
Conclusion
By introducing global pose awareness through a learned "shadow" reference point, this work successfully addresses the inherent limitations of local rotation-invariant descriptors. The resulting framework achieves superior performance across classification and segmentation tasks while maintaining robustness to arbitrary rotations. Future work will explore extending this mechanism to scene-level analysis where multiple objects may have independent orientations.
Summary updated