Few-Shot Unsupervised Implicit Neural Shape Representation Learning with Spatial Adversaries

TL;DR

This paper introduces a regularization term that leverages adversarial samples around the shape to improve the learned SDFs and illustrates the efficacy of the proposed method, highlighting its capacity to improve SDF learning with respect to baselines and the state-of-the-art using synthetic and real data.

Abstract

Implicit Neural Representations have gained prominence as a powerful framework for capturing complex data modalities, encompassing a wide range from 3D shapes to images and audio. Within the realm of 3D shape representation, Neural Signed Distance Functions (SDF) have demonstrated remarkable potential in faithfully encoding intricate shape geometry. However, learning SDFs from sparse 3D point clouds in the absence of ground truth supervision remains a very challenging task. While recent methods rely on smoothness priors to regularize the learning, our method introduces a regularization term that leverages adversarial samples around the shape to improve the learned SDFs. Through extensive experiments and evaluations, we illustrate the efficacy of our proposed method, highlighting its capacity to improve SDF learning with respect to baselines and the state-of-the-art using synthetic and real data.

Figures (10)

• Figure 1: While the training loss (left) is decreasing for both our baseline ma2020neural and our method, the Chamfer distance of reconstructions w.r.t. the GT starts increasing quite early on especially in the sparse input point cloud case for the baseline. This undesirable behaviour is remedied by our adversarial query mining. We report here metrics for unit box normalized meshes, using shape Gargoyle of dataset SRB williams2019deep.
• Figure 2: We learn an implicit shape representation $f_{\theta}$ from a point cloud (blue points) by minimizing the error between projection (through $f_{\theta}$) of spatial queries $q$ (gray points) onto the level set of the field (purple) and the nearest input point $p$. We introduce adversarial queries $q+\hat{\delta}$ to the optimization. They are defined as samples maximizing the loss in the vicinity of original queries.
• Figure 3: ShapeNet shapenet reconstructions from sparse noisy unoriented point clouds.
• Figure 4: Faust Bogo:CVPR:2014 reconstructions from sparse noisy unoriented point clouds.
• Figure 5: 3D Scene zhou2013dense reconstructions from sparse unoriented point clouds.