Introduction
In the field of 3D computer vision, achieving robustness to arbitrary rotations is a critical challenge for applications such as robotics and autonomous driving. Traditional Rotation-Invariant (RI) models often rely on local geometric descriptors—like distances and angles between neighboring points—to ensure that features remain unchanged regardless of the object's orientation. However, this focus on local geometry frequently comes at the cost of "global pose awareness."
A significant drawback identified by the authors is wing-tip feature collapse. In symmetric objects, such as an airplane, the local geometry of the left wing tip is identical to that of the right wing tip. Without a global reference frame, RI models cannot distinguish between these spatially distinct parts, leading to performance degradation in fine-grained tasks like part segmentation.
Shadow-informed Pose Features (SiPF)
To address the loss of global context, this paper introduces the Shadow-informed Pose Feature (SiPF). The core innovation is the introduction of a "shadow"—a globally consistent reference point for every point in the cloud. Unlike absolute coordinates, which change with rotation, the shadow is derived from a shared global rotation matrix that rotates with the object.
By calculating the relationship between a point, its neighbors, and this global "shadow" anchor, the model gains a sense of relative global position without sacrificing rotation invariance. This "shadow" acts as a fixed North Star within the object's own coordinate system, allowing the network to differentiate between symmetric components like left and right wings.
Proposed Framework and Methodology
The authors propose a comprehensive architecture consisting of three main components designed to integrate global awareness into an RI pipeline.
- Task-adaptive Shadow Locating: Rather than using a fixed or arbitrary reference, the model learns the optimal global rotation during training. By modeling the global orientation using a Bingham distribution over unit quaternions, the framework can dynamically adapt the "shadow" location to suit the specific requirements of the task, whether it be classification or segmentation.
- SiPF Extraction: This module generates the descriptors that combine local geometry with the global shadow reference, providing the necessary context for spatial discrimination.
- Rotation-invariant Attention Convolution (RIAttnConv): This novel operator utilizes an attention mechanism to adaptively weight neighboring features based on SiPF descriptors. This allows the network to prioritize features that are most effective at distinguishing spatially distinct but geometrically similar regions.
Experimental Results
The framework was rigorously evaluated across three major 3D benchmarks, demonstrating superior performance compared to existing state-of-the-art RI methods.
- ModelNet40 (Classification): The model achieved a state-of-the-art accuracy of 91.8% using only point coordinates and 92.6% when augmented with surface normals. This outperforms established methods like RI-GCN and VN-DGCNN.
- ShapeNetPart (Segmentation): In the challenging task of part segmentation under arbitrary rotations, the model reached a class mean Intersection-over-Union (C. mIoU) of 82.9%. This result highlights its ability to overcome the "wing-tip feature collapse" and accurately label fine-grained parts.
- ScanObjectNN (Real-world Robustness): On this noisy, real-world dataset, the proposed method achieved 84.0% accuracy, significantly surpassing previous RI benchmarks and proving its practical utility in cluttered environments.
Conclusion
This research presents a significant step forward in rotation-invariant 3D learning by successfully reintroducing global pose awareness. Through the use of Shadow-informed Pose Features and task-adaptive attention, the model effectively mitigates feature collapse in symmetric objects while maintaining strict rotation invariance. While currently focused on object-level analysis, future work aims to extend this "shadow" concept to complex, scene-level environments where multiple objects and increased pose ambiguity present new challenges.
Summary updated