Abstract
Equivariance to permutations and rigid motions is an important inductive bias for various 3D learning problems. Recently it has been shown that the equivariant Tensor Field Network architecture is universal—it can approximate any equivariant function. In this paper, we suggest a much simpler architecture, prove that it enjoys the same universality guarantees, and evaluate its performance on ModelNet40.
Introduction
Deep learning models processing 3D point clouds, such as LiDAR data or CAD models, must respect fundamental geometric symmetries: permutations (invariance to point ordering) and rigid motions (equivariance to rotation and translation). While many equivariant architectures have been proposed, achieving "universality"—the theoretical ability to approximate any continuous equivariant function—often requires complex mathematical machinery. Existing universal models like Tensor Field Networks (TFN) rely on group representation theory, including Wigner D-matrices and spherical harmonics, making them difficult to implement. This work introduces a "semi-lifted" architecture that achieves universality using only standard linear algebra and tensor operations.
Methodology: A Simplified Universal Architecture
The proposed model utilizes a U-Net-like structure composed of three primary parametric layers that operate on tensor representations. Instead of complex irreducible representations, the network uses basic tensor products and contractions.
- Ascending Layers: These layers employ tensor products to map lower-order representations to higher-order ones (e.g., vectors to matrices). This "lifts" the data into a higher-dimensional space to capture intricate geometric relationships.
- Descending Layers: These layers use contractions (generalizations of the matrix trace) to map high-order tensors back to lower orders.
- Linear Layers: Standard layers are used to mix information across different channels.
The complexity of the geometric polynomials the network can represent is determined by the maximal representation order ($K$). The authors prove that for any even $K$, the network can express any equivariant polynomial of degree $\le K$. For instance, $K=6$ is sufficient to compute the eigenvalues of a point cloud's covariance matrix. To enhance practical performance, the architecture incorporates standard components like ReLU activations and K-Nearest Neighbors (KNN) for local spatial context.
Experimental Results
The model was evaluated on the ModelNet40 dataset, consisting of 12,311 CAD models across 40 categories. Testing was conducted under three rotation scenarios: aligned (z/z), trained on aligned but tested on rotated (z/SO(3)), and fully rotated (SO(3)/SO(3)).
Key Findings
- Performance Scaling: Accuracy improved as the tensor order $K$ increased, reaching 81.6% at $K=6$.
- Robust Equivariance: Due to its mathematical design, the model's performance remains consistent regardless of the test data's orientation (e.g., 81.6% for z/z vs. 81.5% for z/SO(3) at $K=6$).
- Parameter Efficiency: The $K=6$ model is highly efficient, utilizing only 370,000 parameters—significantly fewer than other equivariant models like VN-PointNet or VN-DGCNN.
- Ablation Study: The inclusion of non-linearities (ReLU) and local information (KNN) is critical; removing them leads to substantial performance drops (29% and 15.6% respectively).
Conclusion and Future Work
This research demonstrates that universality in rotation-equivariant networks does not require the heavy machinery of physics-inspired models. While a performance gap remains between this simplified universal model and specialized state-of-the-art architectures, the theoretical guarantees and parameter efficiency provide a strong foundation. Future work will focus on optimizing these "semi-lifted" models, exploring higher-order tensors ($K > 6$), and improving channel-mixing layers.
Summary updated