---
title: 'VariGrad: A Feature Architecture for 3D Shapes'
url: https://www.emergentmind.com/topics/varigrad
type: topic
---

# VariGrad: A Feature Architecture for 3D Shapes

Searching arXiv for the primary VariGrad paper and a few directly related baseline methods mentioned in the provided material.
to=arxiv_search tool ,一本道  天天中彩票官网രം json
{"query":"id:2307.03553 OR ti:\"VariGrad: A Novel Feature Vector Architecture for Geometric Deep Learning on Unregistered Data\"","max_results":5,"sort_by":"submittedDate","sort_order":"descending"} สูตรบาคาร่า to=arxiv_search tool  盈立քներ  天天中彩票追号ադարձ json
{"query":"ti:\"PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation\" OR ti:\"Dynamic Graph CNN for Learning on Point Clouds\" OR ti:\"Semi-Supervised Classification with Graph Convolutional Networks\"","max_results":10,"sort_by":"relevance","sort_order":"descending"} to=arxiv_search tool เติมเงินไทยฟรี 山大发 json
{"query":"ti:\"DiffusionNet\" OR ti:\"DeltaConv\" OR ti:\"Field Convolutions for Surface CNNS\"","max_results":10,"sort_by":"relevance","sort_order":"descending"} to=arxiv_search tool 】【。】【”】【json
{"query":"author:Charon AND Trouve varifold OR Kaltenmark varifold OR Besnier varifold","max_results":10,"sort_by":"relevance","sort_order":"descending"} to=arxiv_search tool ＿俺去也 《凤凰大参考  鼎丰json
{"query":"ti:\"Variational Gradient Flow\" OR ti:\"VarGrad: A Low-Variance Gradient Estimator for Variational Inference\"","max_results":10,"sort_by":"relevance","sort_order":"descending"}
I found the primary VariGrad paper and several directly related papers mentioned in the supplied material, including PointNet, DGCNN, Kipf & Welling’s GCN, DiffusionNet, DeltaConv, and VarGrad.
VariGrad is a geometric deep learning layer for *unregistered* 3D geometric data that maps curves and shape graphs to fixed-length feature vectors by differentiating a varifold distance with respect to the vertices of a fixed template shape [2307.03553]. Its central construction is the **varifold gradient**: for an input shape \(\mathcal{C}\) and a template \(\mathcal{T}\), VariGrad computes the gradient of the squared varifold distance \(E(\mathcal{T},\mathcal{C})=\|\mu_T-\mu_C\|^2\) with respect to the template vertices, producing a vector field on the template that can be processed by standard convolutions or graph convolutions [2307.03553]. Because the representation is parameterization independent, the layer can be trained and tested on data independent of the given sampling or parameterization, and the reported experiments demonstrate efficiency, generalizability, and robustness to resampling on 3D face curves and human pose shape graphs [2307.03553].

## 1. Formal definition and geometric representation

VariGrad is built for two input classes: **3D curves** represented as sequences of vertices with edges connecting successive vertices, and **shape graphs** embedded in \(\mathbb{R}^3\) with linear edges between vertices [2307.03553]. Each input is an unregistered shape \(\mathcal{C}=(V,E)\), and the method assumes a fixed template shape \(\mathcal{T}=(V_T,E_T)\) that defines the canonical reference domain on which features will live.

The geometric representation is a **varifold** in ambient space \(\mathbb{R}^3\) with orientation space \(S^2\), hence in \(\mathcal{M}(\mathbb{R}^3\times S^2)\). For each edge \(e\in E\), VariGrad uses its centroid \(c_e\in\mathbb{R}^3\), unit tangent \(u_e\in S^2\), and length \(l_e>0\), and maps the discrete shape graph to the finitely supported varifold
\[
\mu = \sum_{e\in E} l_e\,\delta_{(c_e,u_e)}.
\]
This representation encodes both where edges are and how they are oriented, with weights given by edge lengths [2307.03553].

The associated RKHS kernel is a product of a Gaussian kernel on positions and a Binet kernel on directions:
\[
k\big((x_1,u_1),(x_2,u_2)\big)
=
e^{-a\|x_1-x_2\|^2}\,\langle u_1,u_2\rangle^2,
\]
where \(a>0\) is a scale/balancing parameter controlling spatial sensitivity [2307.03553]. For discrete shape graphs \(\mathcal{C}_1=(V_1,E_1)\) and \(\mathcal{C}_2=(V_2,E_2)\), the induced varifold inner product is
\[
\langle \mu_1,\mu_2\rangle_{\mathrm{Var}}
=
\sum_{e_i\in E_1}\sum_{e_j\in E_2}
e^{-a\|c_i-c_j\|^2}\,\langle u_i,u_j\rangle^2\,l_i l_j,
\]
and the varifold distance is
\[
\|\mu-\nu\|^2
=
\langle \mu,\mu\rangle_{\mathrm{Var}}
+
\langle \nu,\nu\rangle_{\mathrm{Var}}
-
2\langle \mu,\nu\rangle_{\mathrm{Var}}.
\]
The dissimilarity between shape graphs is then \(d_{\mathrm{Var}}(\mathcal{C}_1,\mathcal{C}_2)=\|\mu_1-\mu_2\|\) [2307.03553].

A key property is that the varifold representation is parameterization-independent: resampling, subdividing edges, or reordering vertices does not change the underlying continuous geometric object encoded by centroids, tangents, and lengths in the limit [2307.03553]. This property is the basis for VariGrad’s use on unregistered data.

## 2. The varifold gradient as a feature architecture

The defining computation in VariGrad is the gradient of the squared varifold distance between the template and the input. Let \(\mu_T=\mu(\mathcal{T})\) and \(\mu_C=\mu(\mathcal{C})\). The energy is
\[
E(\mathcal{T},\mathcal{C})=\|\mu_T-\mu_C\|^2.
\]
VariGrad computes
\[
\nabla_{V_T}E(\mathcal{T},\mathcal{C})
=
\left(
\frac{\partial E}{\partial v_{T,1}},\dots,
\frac{\partial E}{\partial v_{T,|V_T|}}
\right),
\]
where each \(v_{T,i}\in\mathbb{R}^3\) is a template vertex [2307.03553]. Since edge centroids, tangents, and lengths depend on template vertex positions, differentiating the energy produces a force-like vector at each template vertex. The paper gives the structural form
\[
\frac{\partial E}{\partial v_{T,i}}
=
\sum_{e\in E_T\ \text{incident to}\ v_{T,i}}
\left(
\frac{\partial E}{\partial c_e}\frac{\partial c_e}{\partial v_{T,i}}
+
\frac{\partial E}{\partial u_e}\frac{\partial u_e}{\partial v_{T,i}}
+
\frac{\partial E}{\partial l_e}\frac{\partial l_e}{\partial v_{T,i}}
\right).
\]

The resulting feature
\[
G(\mathcal{C}) := \nabla_{V_T}E(\mathcal{T},\mathcal{C})
\]
is a tensor of shape \(|V_T|\times 3\), i.e. one 3D vector per template vertex [2307.03553]. This is the central representational move of the method: instead of using a varifold distance only as a loss or similarity score, VariGrad uses the **gradient of the varifold norm** as a descriptor. The feature is defined on the fixed template, so its dimension is constant regardless of how many points or edges the input shape has.

The subsequent processing depends on template topology. For **open curves**, VariGrad uses 1D convolutions along the vertex ordering. For **closed curves**, it uses cyclic padding in 1D convolutions. For **shape graphs** and more general graphs, it uses graph convolutional networks, specifically GCN layers of the Kipf–Welling form [1609.02907]:
\[
H^{(l+1)}=
\sigma\!\left(
\tilde D^{-1/2}\tilde A\tilde D^{-1/2}H^{(l)}W^{(l)}
\right),
\]
with \(H^{(0)}\in\mathbb{R}^{|V_T|\times 3}\) given by the gradient vectors, \(\tilde A\) the template adjacency with self-loops, and \(\tilde D\) the degree matrix [2307.03553]. Edge pooling is then used to reduce dimensionality and aggregate features.

After several convolution and pooling layers, the network produces a global feature vector
\[
f(\mathcal{C})\in\mathbb{R}^N,
\]
which can be passed to an MLP classifier or to a decoder in an autoencoder [2307.03553]. This suggests a canonicalization mechanism: irregular input geometry is first converted into a vector field on a fixed domain, and only then consumed by standard deep architectures.

## 3. Parameterization independence, resampling stability, and invariance boundaries

In this setting, *parameterization* refers to the ordering of vertices, the discretization density, and the specific realization of edges connecting points [2307.03553]. VariGrad’s claim to parameterization independence comes from the varifold layer rather than from a learned invariance. If two discretizations \(\mathcal{C}^{(1)}\) and \(\mathcal{C}^{(2)}\) describe the same underlying curve or graph, then their varifolds \(\mu^{(1)}\) and \(\mu^{(2)}\) are close or equal in varifold space, which implies
\[
\|\mu_T-\mu^{(1)}\|^2 \approx \|\mu_T-\mu^{(2)}\|^2.
\]
Accordingly, the corresponding gradient features satisfy
\[
G(\mathcal{C}_1)\approx G(\mathcal{C}_2)
\]
for reparameterizations of the same geometric object [2307.03553].

The paper’s qualitative invariance experiment makes this point explicit. One shape graph is randomly reparameterized 100 times and passed through trained autoencoders based on PointNet [1612.00593], DGCNN [1801.07829], and VariGrad. The reported outcome is that VariGrad’s outputs cluster tightly around a single reconstruction, whereas PointNet and DGCNN outputs are visibly inconsistent [2307.03553]. The interpretation offered in the paper is that the varifold-based encoding provides sampling invariance and that the gradient-based feature construction preserves it.

A common misconception is that parameterization independence automatically implies full geometric invariance. The paper explicitly notes a boundary condition: **varifold distance is sensitive to absolute position and orientation unless normalization or alignment is performed** [2307.03553]. VariGrad performs well on FAUST shape graphs despite some variability in rigid alignment not seen in training, but the method does not claim rigid invariance as a built-in property.

Another misconception is that VariGrad is merely a registration loss repurposed inside a network. The paper’s novelty claim is narrower and more specific: earlier varifold-based shape analysis used varifold norms as fidelity metrics or loss functions in registration and matching, whereas VariGrad uses the **gradient** of the varifold distance as a **feature extractor** [2307.03553].

## 4. Experimental results on classification, reconstruction, and transfer

The reported evaluation covers two main datasets. **COMA face curves** consist of 20,465 closed 3D curves representing face boundaries extracted from COMA meshes, with 18,506 training curves and 1,959 validation curves, 8 identity labels, and vertex counts between 64 and 96 [2307.03553]. **DFAUST shape graphs** consist of 41,220 3D shape graphs representing human poses extracted from DFAUST meshes, with 35,000 training graphs and 6,220 test graphs, 12 identity labels, and vertex counts between 111 and 356 [2307.03553].

For classification, the pipeline uses one of PointNet, DGCNN, or VariGrad as feature extractor, followed by an MLP classifier with fully connected layers, ReLU, batch normalization, and a final linear layer to \(\mathbb{R}^C\) [2307.03553]. The loss is standard cross-entropy,
\[
\mathcal{L}_{\mathrm{cls}}(\mathcal{C},y)=-\log p_y(\mathcal{C}).
\]

| Architecture | Face curves | Shape graphs |
|---|---:|---:|
| PointNet | 29.36 | 78.08 |
| DGCNN | 62.71 | 83.29 |
| VariGrad | **84.37** | **83.31** |

Training time per batch of 10 samples is also reported. On face curves, VariGrad requires \(\sim 0.045\)s per batch versus \(\sim 0.157\)s for PointNet and \(\sim 0.099\)s for DGCNN. On shape graphs, VariGrad requires \(\sim 0.046\)s versus \(\sim 0.148\)s and \(\sim 0.113\)s, respectively [2307.03553]. The strongest gain appears on the sparser face-curve regime.

For reconstruction, the encoder is again PointNet, DGCNN, or VariGrad, and the decoder is a fully connected network outputting vertex positions for a target template-like graph [2307.03553]. The reconstruction loss is the squared varifold norm,
\[
\mathcal{L}_{\mathrm{rec}}(\mathcal{C},\widehat{\mathcal{C}})
=
\|\mu(\mathcal{C})-\mu(\widehat{\mathcal{C}})\|^2,
\]
which is differentiable and parameterization-independent.

| Architecture | Face curves | Shape graphs |
|---|---:|---:|
| PointNet | 0.0053 | 0.0039 |
| DGCNN | 0.0010 | 0.0015 |
| VariGrad | **0.0004** | **0.0005** |

For reconstruction training time per batch of 10 samples, VariGrad is again the fastest: face curves \(\sim 0.073\)s versus \(\sim 0.149\)s and \(\sim 0.113\)s; shape graphs \(\sim 0.072\)s versus \(\sim 0.174\)s and \(\sim 0.133\)s [2307.03553].

The transfer and invariance experiments use two additional test settings: **reparameterized DFAUST**, consisting of randomly resampled and reparameterized versions of DFAUST test graphs with identity labels, and **FAUST shape graphs**, 90 graphs extracted from FAUST with different poses and some variability in rigid alignment [2307.03553].

| Architecture | Reparam DFAUST: Acc (%) | Reparam DFAUST: mean error |
|---|---:|---:|
| PointNet | 49.98 | 0.0040 |
| DGCNN | 50.63 | 0.0049 |
| VariGrad | **80.91** | **0.0005** |

| Architecture | FAUST: mean error |
|---|---:|
| PointNet | 0.0061 |
| DGCNN | 0.0064 |
| VariGrad | **0.0040** |

These results support the paper’s main empirical thesis: a feature architecture grounded in parameterization-independent varifolds and evaluated on a fixed template can maintain performance under sampling changes that substantially degrade point-based baselines [2307.03553].

## 5. Position within geometric deep learning and nearby nomenclature

VariGrad is positioned against point-based architectures such as PointNet [1612.00593] and DGCNN [1801.07829]. PointNet processes each point independently through shared MLPs and aggregates pointwise features using symmetric functions; DGCNN builds a dynamic \(k\)-NN graph on the point cloud and applies edge convolutions and pooling [2307.03553]. The paper identifies four limitations of these baselines for curves and shape graphs: they treat vertices as unordered points, lose explicit edge and tangent information, depend on learned neighborhoods induced by the sampled points rather than on known connectivity, and are sensitive to sampling density and point ordering [2307.03553]. VariGrad’s counter-design is to use edge information directly, anchor features on a fixed template, and apply graph convolutions on the template connectivity rather than on a dynamically rebuilt neighborhood graph.

The paper also distinguishes its target domain from surface-centric geometric networks. Methods such as DiffusionNet [2012.00888] use surface-specific operators such as Laplace–Beltrami-type constructions, whereas VariGrad specifically targets **3D curves and shape graphs** and does not require surface-related differential operators [2307.03553]. This is not a claim of broader scope; it is a domain specialization.

Relative to older varifold-based shape analysis, the novelty is not the use of varifold norms per se but the use of the **gradient field of the varifold distance** as a learnable representation [2307.03553]. The gradient can be interpreted as the optimal local deformation direction of each template vertex under the chosen metric, but the method stops short of explicit deformation optimization in the registration sense.

The name should also be distinguished from unrelated or only partially related arXiv nomenclature. "Deep Generative Learning via Variational Gradient Flow" [1901.08469] concerns variational gradient flow on probability spaces for generative modeling, and "VarGrad: A Low-Variance Gradient Estimator for Variational Inference" [2010.10436] concerns an unbiased score-function estimator for ELBO gradients. Those works share lexical proximity, not the geometric deep learning construction introduced in "VariGrad: A Novel Feature Vector Architecture for Geometric Deep Learning on Unregistered Data" [2307.03553].

## 6. Implementation, limitations, and stated extensions

The implementation described in the paper is in **PyTorch**, with code and pretrained models made available on GitHub [2307.03553]. The template \(\mathcal{T}\) is selected as a **random shape** from the training set for each dataset, and this template supplies both the graph structure for convolutions and the coordinate system on which the gradient feature is defined. The varifold energy is differentiated with respect to template vertices via automatic differentiation by defining \(E(\mathcal{T},\mathcal{C})\) and calling `.backward()` [2307.03553].

The discrete varifold inner product scales as
\[
\sum_{e_i\in E_T}\sum_{e_j\in E_C} k(e_i,e_j)\,l_i l_j,
\]
hence has complexity \(O(|E_T|\cdot |E_C|)\) [2307.03553]. The paper notes that this is manageable for the curve and graph sizes studied, but it may become computationally heavy for very large graphs or dense surfaces. Even so, the reported training times per batch are lower than those of PointNet and DGCNN, which the paper attributes to fixed template size, fewer trainable parameters, and efficient graph convolutions on the template [2307.03553].

Several limitations are stated or implied. VariGrad assumes that data are given as **vertices and edges**, so it may require domain-specific preprocessing such as landmark extraction, geodesic computation on surfaces, or concatenation of geodesic curves into shape graphs [2307.03553]. The use of a **single template** also introduces a structural bias: if a class exhibits high topological variability, template selection becomes difficult, and the paper explicitly notes that hand sketches may require careful template selection [2307.03553]. In addition, the current experiments focus on curves and shape graphs rather than surfaces or raw point clouds.

The paper outlines three immediate directions for extension. The first is **hand sketches**, where within-class topological variation is large [2307.03553]. The second is **surfaces and point clouds**, since varifold representations naturally generalize to them even though high-resolution efficiency and memory remain open concerns [2307.03553]. The third is integration with more complex generative architectures such as **VAE** and **GAN**, motivated by prior use of varifold-based metrics in mesh-invariant generative models [2307.03553].

Taken together, these elements define VariGrad as a template-based, varifold-gradient feature architecture for unregistered 3D curves and shape graphs: it converts each input into a parameterization-independent vector field on a fixed template, processes that field with topology-aware convolutions, and yields fixed-length descriptors that have shown strong performance and strong robustness to resampling on the reported benchmarks [2307.03553].

Source: https://www.emergentmind.com/topics/varigrad