Learning Smooth Neural Functions via Lipschitz Regularization

Hsueh-Ti Derek Liu,Francis Williams,Alec Jacobson,Sanja Fidler,Or Litany

DOI: https://doi.org/10.48550/arXiv.2202.08345

2022-05-11

Abstract:Neural implicit fields have recently emerged as a useful representation for 3D shapes. These fields are commonly represented as neural networks which map latent descriptors and 3D coordinates to implicit function values. The latent descriptor of a neural field acts as a deformation handle for the 3D shape it represents. Thus, smoothness with respect to this descriptor is paramount for performing shape-editing operations. In this work, we introduce a novel regularization designed to encourage smooth latent spaces in neural fields by penalizing the upper bound on the field's Lipschitz constant. Compared with prior Lipschitz regularized networks, ours is computationally fast, can be implemented in four lines of code, and requires minimal hyperparameter tuning for geometric applications. We demonstrate the effectiveness of our approach on shape interpolation and extrapolation as well as partial shape reconstruction from 3D point clouds, showing both qualitative and quantitative improvements over existing state-of-the-art and non-regularized baselines.

Computer Vision and Pattern Recognition,Graphics

What problem does this paper attempt to address?

The paper attempts to address the issue of achieving smoothness with respect to latent parameters in Neural Implicit Fields. Specifically, the authors found that existing methods fail to ensure the smoothness of network outputs outside the training set, especially in tasks such as shape interpolation and extrapolation. To overcome this limitation, they propose a novel regularization technique that encourages smoothness by penalizing the upper bound of the Lipschitz constant in the neural field's latent space. This method is not only simple and feasible but also effectively enhances smoothness and robustness in geometric applications. The authors demonstrate the effectiveness of their method in tasks such as shape interpolation, extrapolation, and shape completion from partial point clouds, outperforming existing state-of-the-art techniques and non-regularized baselines both qualitatively and quantitatively. Additionally, the robustness of the method under adversarial attacks is significantly improved. In summary, this study aims to improve the smoothness of neural field representations by introducing Lipschitz regularization, making operations like shape editing more stable and reliable.

Learning Smooth Neural Functions via Lipschitz Regularization

Implicit Geometric Regularization for Learning Shapes

Augmenting Implicit Neural Shape Representations with Explicit Deformation Fields

1-Lipschitz Neural Distance Fields

Regularization Neural Networks Via Constrained Virtual Movement Field

Markov-Lipschitz Deep Learning

StEik: Stabilizing the Optimization of Neural Signed Distance Functions and Finer Shape Representation

Neural Octahedral Field: Octahedral prior for simultaneous smoothing and sharp edge regularization

Learning with Hyperspherical Uniformity

LISR: Learning Linear 3D Implicit Surface Representation Using Compactly Supported Radial Basis Functions

Strained virtual movement field

Deep Implicit Moving Least-Squares Functions for 3D Reconstruction

Learning smooth functions in high dimensions: from sparse polynomials to deep neural networks

Learning continuous shape priors from sparse data with neural implicit functions

Learning Manifold Patch-Based Representations of Man-Made Shapes

Smoothing Neural Network for L 0 Regularized Optimization Problem with General Convex Constraints

Neural Representation of Shape-Dependent Laplacian Eigenfunctions

Controlled Learning of Pointwise Nonlinearities in Neural-Network-Like Architectures

Learning Sparse Visual Representations with Leaky Capped Norm Regularizers

Implicit Neural Head Synthesis via Controllable Local Deformation Fields

A Unified Algebraic Perspective on Lipschitz Neural Networks