Abstract
This paper introduces SE(3) Separable Point Convolution (SPConv), a novel framework designed to make 3D point cloud analysis robust to rotations and translations without the prohibitive computational costs typically associated with 6D convolutions. By factorizing complex operations into spatial and rotational components and employing a Group Attentive (GA) pooling mechanism, the proposed method achieves state-of-the-art performance in 3D shape classification and alignment.
Introduction
In 2D computer vision, translation equivariance is a cornerstone of CNN success. However, 3D data requires SE(3) equivariance, accounting for both translations and arbitrary rotations. Existing methods often choose between invariant descriptors, which discard structural detail, or equivariant networks, which convolve over a 6D space (3D position + 3D rotation). The latter is often computationally impractical for large-scale scans. This work bridges the gap by providing an expressive yet efficient equivariant network.
Methodology: SE(3) Separable Point Convolution
The core innovation is the SE(3) Separable Point Convolution (SPConv). Inspired by depthwise separable convolutions in 2D networks, the authors decompose the 6D convolution into two sequential steps:
- SE(3) Point Convolution: Aggregates spatial information in 3D Euclidean space ($R^3$).
- SE(3) Group Convolution: Aggregates information across the rotation group ($SO(3)$).
This factorization reduces computational complexity from a multiplicative relationship ($O(K_p \cdot K_g)$) to an additive one ($O(K_p + K_g)$), where $K_p$ and $K_g$ represent spatial and rotational kernel sizes, respectively.
Group Attentive (GA) Pooling
Standard pooling operations can disrupt rotation-equivariant feature patterns. To address this, the authors introduce Group Attentive (GA) Pooling. This mechanism uses attention to learn an "intrinsic local frame," assigning weights to different rotations to align features into a canonical orientation before fusion into a rotation-invariant representation.
Experimental Evaluation
The model was evaluated on two primary benchmarks:
- Rotated ModelNet40: Testing robustness to orientation in shape classification.
- 3DMatch: Evaluating real-world shape registration and scan alignment.
Key Results
- Pose Estimation: On ModelNet40 (airplane category), the model achieved a median angular error of 1.11°, significantly better than the KPConv baseline (8.06°).
- Classification: In the rotated ModelNet40 setting, the model reached 88.3% accuracy, outperforming PointNet++ (85.0%) and KPConv (86.7%).
- 3D Registration: The model set a new state-of-the-art on 3DMatch with an average recall of 97.6%.
Discussion and Insights
The mathematical factorization of SE(3) convolution makes equivariant networks viable for standard hardware. The network effectively treats pose estimation as a detection task over discrete rotation "anchors." Interestingly, the attention mechanism shows high confidence on asymmetric objects but naturally reflects ambiguity on symmetric shapes, confirming its geometric awareness.
While highly efficient, the model currently uses a discrete rotation group (icosahedral), which is an approximation of continuous space. Future work may explore more continuous representations or further pruning to reduce the overhead of maintaining multiple rotation states.
Summary updated