Neural Unsigned Distance Fields for Implicit Function Learning

Chibane, Julian; Mir, Aymen; Pons-Moll, Gerard

doi:10.48550/arxiv.2010.13938

Cited by 10 publications

(22 citation statements)

References 48 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We present a simple approach to PC extraction from DDFs, though it cannot guarantee uniform sampling over the shape. Analogous to prior work [5], we recall that q(p, v)…”

Section: Single-image 3d Reconstructionmentioning

confidence: 99%

“…Implicit Shape Representations Our work is most similar to distance field representations of shape, which have a long history in computer vision [60], most recently culminating in signed and unsigned distance fields (S/UDFs) [5,51,74]. In comparison to explicit representations, implicit shapes can capture arbitrary topologies with high fidelity.…”

Section: Related Workmentioning

confidence: 99%

“…2 When the surface normals exist, v * is closely related to them: v * (p) = −n(p, v * (p)), in the notation of Property II. Recent work has highlighted the utility of UDFs over SDFs [5,74]; in the case of DDFs, extracting a UDF or v * may provide useful auxiliary information for some tasks.…”

Section: Udf Extractionmentioning

confidence: 99%

See 2 more Smart Citations

Representing 3D Shapes with Probabilistic Directed Distance Fields

Aumentado-Armstrong¹,

Tsogkas²,

Dickinson³

et al. 2021

Preprint

View full text Add to dashboard Cite

Differentiable rendering is an essential operation in modern vision, allowing inverse graphics approaches to 3D understanding to be utilized in modern machine learning frameworks. Explicit shape representations (voxels, point clouds, or meshes), while relatively easily rendered, often suffer from limited geometric fidelity or topological constraints. On the other hand, implicit representations (occupancy, distance, or radiance fields) preserve greater fidelity, but suffer from complex or inefficient rendering processes, limiting scalability. In this work, we endeavour to address both shortcomings with a novel shape representation that allows fast differentiable rendering within an implicit architecture. Building on implicit distance representations, we define Directed Distance Fields (DDFs), which map an oriented point (position and direction) to surface visibility and depth. Such a field can render a depth map with a single forward pass per pixel, enable differential surface geometry extraction (e.g., surface normals and curvatures) via network derivatives, be easily composed, and permit extraction of classical unsigned distance fields. Using probabilistic DDFs (PDDFs), we show how to model inherent discontinuities in the underlying field. Finally, we apply our method to fitting single shapes, unpaired 3D-aware generative image modelling, and single-image 3D reconstruction tasks, showcasing strong performance with simple architectural components via the versatility of our representation.

show abstract

“…We present a simple approach to PC extraction from DDFs, though it cannot guarantee uniform sampling over the shape. Analogous to prior work [5], we recall that q(p, v)…”

Section: Single-image 3d Reconstructionmentioning

confidence: 99%

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Representing 3D Shapes with Probabilistic Directed Distance Fields

Aumentado-Armstrong¹,

Tsogkas²,

Dickinson³

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…Deep neural networks for computational science While deep neural networks were originally designed for tackling problems of regression or classification, their use has now been extended to deal with different sorts of problems in computational sciences and applications. Deep neural networks can be trained for computing the SDF to a given point-cloud or triangle mesh [1,2,3,7,11,9,22,21,24]. All these methods differ in how they model the loss function or the network architecture.…”

Section: Related Workmentioning

confidence: 99%

“…It is possible to add other constraints to the loss (7) or the loss (8). For example, if necessary, one could prevent eventual extra zero level-sets away from the surface by adding a term such as…”

Section: Loss Functionsmentioning

confidence: 99%