Deep Geometric Prior for Surface Reconstruction

Francis Williams1 , Teseo Schneider1 , Claudio Silva 1 , Denis Zorin1 , Joan Bruna1 , and Daniele Panozzo1
1
New York University
francis.williams@nyu.edu, teseo.schneider@nyu.edu, csilva@nyu.edu, dzorin@cs.nyu.edu,
bruna@cims.nyu.edu, panozzo@nyu.edu
arXiv:1811.10943v2 [cs.CV] 5 Apr 2019

Abstract retains critical surface features and approximates the sam-

pled surface well, is a pervasive and challenging problem.
The reconstruction of a discrete surface from a point Different approaches have been proposed, mostly
cloud is a fundamental geometry processing problem that grouped into several categories: (1) using the points to de-
has been studied for decades, with many methods devel- fine a volumetric scalar function whose 0 level-set corre-
oped. We propose the use of a deep neural network as a sponds to the desired surface, (2) attempt to ”connect the
geometric prior for surface reconstruction. Specifically, we dots” in a globally consistent way to create a mesh, (3) fit a
overfit a neural network representing a local chart parame- set of primitive shapes so that the boundary of their union
terization to part of an input point cloud using the Wasser- is close to the point cloud, and (4) fit a set of patches to the
stein distance as a measure of approximation. By jointly point cloud approximating the surface.
fitting many such networks to overlapping parts of the point
cloud, while enforcing a consistency condition, we compute We propose a novel method, based, on the one hand, on
a manifold atlas. By sampling this atlas, we can produce a constructing a manifold atlas commonly used in differen-
dense reconstruction of the surface approximating the input tial geometry to define a surface, and, on the other hand, on
cloud. The entire procedure does not require any training observed remarkable properties of deep image priors [37],
data or explicit regularization, yet, we show that it is able using an overfitted neural network for interpolation. We
to perform remarkably well: not introducing typical overfit- define a set of 2D parametrizations, each one mapping a
ting artifacts, and approximating sharp features closely at square in parametric space to a region of a surface, ensur-
the same time. We experimentally show that this geomet- ing consistency between neighbouring patches. This prob-
ric prior produces good results for both man-made objects lem is inherently ambiguous: there are many possible valid
containing sharp features and smoother organic objects, as parametrizations, and only a small subset will correspond to
well as noisy inputs. We compare our method with a number a faithful representation of the underlying surface. We com-
of well-known reconstruction methods on a standard sur- pute each parametrization by overfiting a network to a part
face reconstruction benchmark. of the point cloud, while enforcing consistency conditions
between different patches. We observe that the result is a
reconstruction which is superior quantitatively and qualita-
1. Introduction tively to commonly used surface reconstruction methods.
We use the Wasserstein distance as a training loss, which
3D geometry is commonly acquired in the form of col-
is naturally robust to outliers, and has the added advantage
lections of (possibly incomplete) range images (laser scan-
of providing us explicit correspondences between the para-
ning, structured light, etc) or measurements of more com-
metric domain coordinates and the fitted points, allowing us
plex structure (LIDAR). Unordered set of points (point
to explicitly measure, and thus minimize, the disagreement
clouds) is a commonly used representation of combined
between neighbouring patches covering a surface.
registered results of scanning objects or scenes. Point
clouds can be obtained in other ways, e.g., by sampling an We use a standard shape reconstruction benchmark to
implicit surface using ray casting. Computing a continu- compare our method with 12 competing methods, showing
ous representation of a surface from the discrete point cloud that, despite the conceptual simplicity of our algorithm, our
(e.g., a polygonal mesh, an implicit surface, or a set of pa- reconstructions are superior in terms of quantitative and vi-
rameteric patches) in way that is robust to noise, and yet sual quality.

1
2. Related work nomenclature. They separate techniques into four main cat-
egories: indicator function, point-set surfaces, multi-level
Geometric Deep Learning A variety of architectures partition of unity, and scattered point meshing.
were proposed for geometric applications. A few work with
Indicator function techniques define a scalar function in
point clouds as input; in most cases however, these methods
space that can be used for testing if a given point is in-
are designed for classification or segmentation tasks. One
side or outside the surface. There are multiple ways to
of the first examples are PointNet [33] and PointNet++ [34]
define such a function from which a surface is generated
originally designed for classification and segmentation, us-
by isocontouring. Poisson surface reconstruction (Pois-
ing a set-based neural architecture [39, 36]. PCPNet [14]
son) [21] inverts the gradient operator by solving a Pois-
is version of PointNet architecture, for estimation of local
son equation to define the indicator function. Fourier sur-
shape properties. A number of learning architectures for
face reconstruction (Fourier) [20] represents the indicator
3D geometry work with voxel data converting input point
function in the Fourier basis, while Wavelet surface recon-
clouds to this form (e.g., [38]). The closest problems these
struction (Wavelet) [27] employs a Haar or a Daubechies
types of networks solve is shape completion and point cloud
(4-tap) basis. Screened Poisson surface reconstruction
upsampling.
(Screened) [22] is an extension of [21] that incorporates
Shape completion is considered, e.g., in [9], where vol- point constraints to avoid over smoothing of the earlier tech-
umetric CNN is used to predict a very course shape com- nique. This technique is not considered in [5], but is con-
pletion, which is then refined using data-driven local shape sidered by us.
synthesis on small volumetric patches. [16], follows a
Point-set surfaces [2] define a projection operator that
somewhat similar approach, combining multiview and vol-
moves points in space to a point on the surface, where the
umetric low-resolution global data at a first stage, and using
surface is defined to be the collection of stationary points of
a volumetric network to synthesize local patches to obtain
the projection operator. Providing a definition of the projec-
higher resolution. Neither of these methods aims to achieve
tion operators are beyond the scope of our paper (see [5]).
high-quality surface reconstruction.
In our experiments, we have used simple point set surfaces
PU-Net, described in [41], is to the best of our knowl- (SPSS) [1], implicit moving least squares (IMLS) [24], and
edge, the only learning-based work addressing point cloud algebraic point set surfaces (APSS) [13].
upsampling directly. The method proceeds by splitting in-
Edge-Aware Point Set Resampling (EAR) [18] (also not
put shapes into patches and learning hierarchical features
considered in [5], but considered by us) works by first com-
using PointNet++ type architecture. Then feature aggrega-
puting reliable normals away from potential singularities,
tion and expansion is used to perform point set expansion in
followed by a resampling step with a novel bilateral filter
feature space, followed by the (upsampled) point set recon-
towards surface singularities. Reconstruction can be done
struction.
using different techniques on the resulting augmented point
In contrast to other methods, the untrained networks in set with normals.
our method take parametric coordinates in square paramet-
Multi-level Partition of Unity defines an implicit func-
ric domains as inputs and produce surface points as output.
tion by integrating weight function of a set of input points.
An important exception is the recent work [12] defining an
The original approach of [29] (MPU) uses linear functions
architecture, AtlasNet, in which the decoder part is simi-
as low-order implicits, while [28] (MPUSm) defines differ-
lar to ours, but with some important distinctions discussed
ential operators directly on the MPU function. The method
in Section 3. Finally, [4] studied the ability of neural net-
of [31] (RBF) uses compactly-supported radial basis func-
works to approximate low-dimensional manifolds, showing
tions.
that even two-layer ReLU networks have remarkable abil-
ity to encode smooth structures with near-optimal number Scattered Point Meshing [30] (Scattered) grows
of parameters. In our setting, we rely on overparametrisa- weighted spheres around points in order to determine the
tion and leverage the implicit optimization bias of gradient connectivity in the output triangle mesh.
descent. The work in [32] uses a manifold-based approach to a
direct construction of a global parametrization from a set
of range images (a point cloud, or any other surface repre-
Surface Reconstruction is an established research area sentation, can be converted to such a set by projection to
dating back at least to the early 1990s (e.g, [17]); [6] is multiple planes). It uses range images as charts with pro-
a recent comprehensive survey of the area. We focus our jections as chart maps; our method computes chart maps by
discussion on the techniques that we use for comparison, fitting. [40] jointly optimizes for connectivity and geome-
which are a superset of those included in the surface re- try to produce a single mesh for an entire input point cloud.
construction benchmark of Berger et al [5]. Berger tested In contrast, our method produces a global chart map using
10 different techniques in their paper; we will follow their only a local optimization procedure.
 X1 We propose to approximate ϕ using a neural network,
X2 φ(·; θp ) : V → R3 , where θp is a vector of parameters, that
we fit so that the image of φ approximates Up ∩ S. For that
~
L(µ )
purpose, we first consider a sample {v1 , . . . , vn } of n =
1
X X3 |Xp | points in V using a Poisson disk distribution, and the
~ associated Earth Movers or 2-Wasserstein Distance (EMD)
L(µ 2
)
between {φ(vi ; θp ) : i = 1 . . . n} and Xp = {x1 . . . , xn }:
Á(v; µ1)
~
L(µ )
X
3
L(θ) = inf kφ(vi ; θ) − xπ(i) k2 . (1)
Á(v; µ2) π∈Πn
i≤n

v Á(v; µ3) where Πn is the set of permutations of n points. The

computation of the EMD in (1) requires solving a linear
Figure 1. Overview of our proposed surface reconstruction archi- assignment problem, which can be done in O(n3 ) using,
tecture. An input point cloud X is split into overlapping patches for instance, the Hungarian algorithm [25]. Since this is
Xp , where a local chart ϕp (v) = φ(v, θp ) is obtained via the min-
prohibitive for typical values of n, we rely instead on the
imisation of a Wasserstein loss L̃(θ) (2). The local charts are made
Sinkhorn regularized distance [8]:
globally consistent thanks to the correspondences arising during
the Wasserstein loss minimization (Section 3.2) X
. L̃(θ) = min Pi,j kφ(vj ; θ) − xi k2 − λ−1 H(P ) (2)
P ∈Pn
i,j≤n

Deep Image Prior Our approach is inspired, in part, by where Pn is the set of n × n bi-stochastic
P matrices and
the deep image prior. [37] demonstrated that an untrained H(P ) is the entropy, H(P ) = − i,j Pij log Pij . This
deep network can be overfitted to input data producing distance provably approximates the Wasserstein metric as
a remarkably high-quality upsampling and even hole fill- λ → ∞ and can be computed in near-linear time [3]. Fig-
ing without training, with the convolutional neural network ure 13 in the supplemental material shows the effect of vary-
structure acting as a regularizer. Our approach to surface re- ing the regularization parameter λ on the results.
construction is similar, in that we use untrained networks to We choose φ to be a MLP with the half-rectified activa-
represent individual chart embedding functions. However, tion function:
an important difference is that our loss function measures
geometric similarity. φ(v; θ) = θK ReLU(θK−1 ReLU . . . ReLU(θ1 v)),

where θi , i = 1 . . . K, are per-layer weight matrices. This

3. Method choice of activation function implies that we are fitting a
Our method for surface reconstruction uses a set of deep piece-wise linear approximation to Xp . We choose to over-
ReLU networks to obtain local charts or parametrizations parametrize the network such that the total number of train-
(Section 3.1). These parameterizations are then made con- able parameters T = dim(θ1 ) + · · · + dim(θK ), where dim
sistent with each other on the parts where they overlap (Sec- refers to the total number of entries in the matrix, satisfies
tion 3.2). The networks are trained using the 2-Wasserstein T  3n, which is the number of constraints.
distance as a loss function. The overall architecture of our Under such overparametrized conditions, one verifies
technique is illustrated in Figure 1. that gradient-descent converges to zero-loss in polynomial
In the following, we denote by S a smooth surface (pos- time for least-squares regression tasks [10]. By observing
sibly with a boundary) in R3 . The goal of surface recon- that
struction is to estimate S from a possibly noisy point cloud min L̃(θ) = min min L̃(θ; P ), with
θ P ∈Pn θ
X = {xi = yi + wi ; yi ∈ S; i ≤ N }, where wi models the X
acquisition noise. L̃(θ; P ) = Pi,j kφ(vj ; θ) − xi k2 − λ−1 H(P ),
i,j≤n
3.1. Local Parametrization Model
and that L̃(θ; P ) is convex with respect to P , it follows that
Let us first consider the reconstruction of a given neigh- gradient-descent can find global minimsers of L̃ in polyno-
borhood of S around a point p, denoted as Up , from the cor- mial time.
responding point cloud Xp = X ∩ Up . If Up is sufficiently As λ → ∞, the entropic constraint disappears, which
small, from the implicit function theorem, one can charac- implies that by setting P to any arbitrary permutation ma-
terize Up ∩ S as the image of the open square V = (0, 1)2 trix Π, we can still obtain zero loss (minθ L̃(θ, Π) = 0). In
by a differentiable mapping ϕ : V → Up . other words, the model has enough capacity to produce any
arbitrary correspondence between the points {vi }i and the B(q; c̃r) consists of multiple connected components, possi-
targets {xi }i in the limit λ → ∞. A priori, this is an unde- bly of nontrivial genus. We use the heuristic described be-
sirable property of our model, since it would allow highly low to filter out points we expect to be on a different sheet
oscillatory and irregular surface reconstructions. However, from the ball center q.
our experiments (Section 4) reveal that the gradient-descent To ensure consistency as defined above, we fit the charts
optimization of L̃ remarkably biases the model towards so- in two phases. In the first phase, we locally fit each chart
lutions with no apparent oscillation. This implicit regular- to its associated points. In the second phase, we compute a
isation property is reminiscent of similar phenomena en- joint fitting of all pairs of overlapping charts.
joyed by gradient-descent in logistic regression [35, 15] or Let (θp , πp ) and (θq , πq ) denote the parameters and per-
matrix factorization [26]. In our context, gradient-descent mutations of the patches p and q respectively at some iter-
appears to be favoring solutions with small complexity, ation of the optimization. We compute the first local fitting
measured for instance in terms of the Lipschitz constant of phase as
φθ , without the need of explicit regularization. We leave the X
theoretical analysis of such inductive bias for future work min inf kφ(vi ; θp ) − xπp (i) k2 . (3)
θp πp
(Section 5). Note that, in practice, we set λ to a large value, i≤|Xp |
which may have an extra regularizing effect.
We define the set of indices of parametric points in chart
3.2. Building a Global Atlas p of the intersection Xp,q as
Section 3.1 described a procedure to obtain a local chart Tpq = {i|xπp (i) ∈ Xp,q },
around a point p ∈ S, with parametric domain Vp and its
associated fitted parametrization ϕp = φ(·; θp∗ ) : Vp → where Tqp is the corresponding set in chart q. The map
R3 . In this section, we describe how to construct an atlas between indices of corresponding parametric points in two
{(Vq , ϕq ); q ∈ Q} by appropriately selecting a set of an- patches is given by: πp→q := πq ◦ πp−1 : Tpq → Tqp .
chor points Q and by ensuring consistency between charts. Equipped with this correspondence, we compute the sec-
ond joint fitting phase between all patch pairs as:
Consistency. To define atlas consistency more precisely, X
min inf kφ(vi ; θp ) − φ(vπp→q (i) ; θq )k2 . (4)
we need to separate the notions of parametric coordinate- θp ,θq πp ,πq
i∈Tp;q
assignment and surface approximation, since the local chart
functions ϕp define both. We say that two charts p and q Observe that by the triangle inequality,
overlap, if Xp,q = Xp ∩ Xq 6= ∅. Each discrete chart ϕα is
equipped with a permutation πα , assigning indices of points kφ(vi ; θp ) − φ(vπp→q (i) ; θq )k ≤kφ(vi ; θp ) − xπp (i) k+
in Xp to indices of parametric positions in Vp . Two overlap- kφ(vπp→q (i) ; θq ) − xπp (i) k.
ping charts p and q are consistent on the surface samples
if Therefore, the joint fitting term (4) is bounded by the sum
of two separate fitting terms (3) for each patch (note that
ϕq (vπq−1 (i) ; θq ) = ϕp (vπp−1 (i) ; θp ), ∀xi ∈ Xp ∩ Xq xπp (i) = xπq (πp→q (i)) ). Consistent transitions are thus en-
forced by the original Wasserstein loss if the charts are in-
i.e., for any point in the patch overlap, the values of the two terpolating, i.e. φ(vi ; θp ) = xπp (i) for all i, p. However,
chart maps at corresponding parametric values coincide. If in presence of noisy point clouds, the joint fitting phase en-
all chart maps are interpolating, then consistency is guaran- ables a smooth transition between local charts without re-
teed by construction, but this is in general not the case. We quiring an exact fit through the noisy samples.
enforce consistency explicitly by minimizing a consistency If the Sinkhorn distance is used instead of the EMD, then
loss (4). we project the stochastic matrices Pq , Pq0 to the nearest per-
mutation matrix by setting to one the maximum entry in
Constructing the Atlas. We construct the set of patch each row.
centers Q using Poisson disk sampling [7] of X , with a
specified radius, r. For each q ∈ Q, we first extract a neigh- Filtering Sample Sets Xr . In our experiments, we choose
borhood Xq by intersecting a ball of radius cr centered at the ball radius to be sufficiently small to avoid most of the
q with X (X ∩ B(q; cr)), where c > 1 is another hyper- artifacts related to fitting patches to separate sheets of the
parameter. To reduce boundary effects, we consider a larger surface-intersection with B(q; cr). The radius can be easily
radius c̃ > c and use Xeq := X ∩ B(q; c̃r) to fit the local chosen adaptively, although at a significant computational
chart for Xq . In general, the intersection of S with the ball cost, by replacing a ball by several smaller balls whenever
 Sinkhorn loss, we use a regularization parameter, λ = 1000.
For the networks, φ, we use an MLP with fully connected
layer sizes: (2, 128, 256, 512, 512, 3) and ReLU activa-
tions. Our reference implementation is available at https:
//github.com/fwilliams/deep-geometric-prior.

Single Patch Fitting. Our first experiment shows the be-

haviour of a single-patch network overfitted on a complex
Input Fitted surface point cloud (Figure 2 left). Our result is a tight fit to the
Figure 2. Single patch fitting with uv-mapping illustrated with a point cloud. An important side effect of our construction is
checkerboard texture. an explicit local surface parametrization, which can be used
to compute surface curvature, normals, or to apply an image
onto the surface as a texture map (Figure 2 right).
Figure 3 shows the evolution of the fitting and of the pa-
rameterisation ϕθ as the optimization of L̃(θ) progresses.
We observe that the optimization path follows a trajectory
where φθ does not exhibit distortions, supporting the hy-
pothesis that gradient descent biases towards solutions with
25 75 100 low complexity.

3.2e−1

1.0e−1
Global Consistency. As described in Section 3.2, recon-
3.2e−2 structing an entire surface from local charts requires consis-
Loss

1.0e−2

3.2e−3
tent transitions, leading to the formulation in (3) and (4)
1.0e−3
0 100 200 300 400 500
which reinforces consistency across overlapping patches.
Iteration

150 500 Convergence

Figure 5 illustrates the effect of adding the extra consistency
terms. We verify that these terms significantly improve the
Figure 3. Evolution of single-patch reconstruction as the loss L̃(θ) consistency.
is minimised. The jumps in the loss are a side-effect of the adaptive
gradient descent scheme ADAM.
Surface Reconstruction Benchmark. To evaluate quan-
titatively the performance of our complete reconstruction
the quality of the fit is bad. Instead, we use a cheaper heuris- pipeline, we use the setup proposed by [5], using the first
tic to eliminate points from each set that are likely to be on set of range scans for each of the 5 benchmark models. Fig-
a different sheet: We assume that the input point cloud X ure 4 shows the results (and percentage error) of our method
also contains normals. If normals are unavailable, they can on the five models of the benchmark. We compare our re-
be easily estimated using standard local fitting techniques. sults against the baseline methods described in Section 2,
We then discard all vertices in each Xp whose normals form and use the following metrics to evaluate the quality of the
an angle greater than a fixed threshold α with respect to the reconstruction: Let X denote the input point cloud. From
normal at the center. In all our experiments (Section 4) we the ground-truth surface S and the reconstruction S,b we ob-
used r = 2.5% of the bounding box diagonal enclosing the tain two dense point clouds that we denote respectively by
point cloud, c = 1.5 c̃ = 1.5, α = 100 degrees. Y and Y.b We consider

dinp→rec (i) = minybj ∈Yb kxi − ybj k , i ≤ |X |,

4. Experiments
drec→GT (j) = minyk ∈Y kb
yj − yk k , j ≤ |Y|.
b (5)
Experimental Setup. We run our experiments on a com-
puting node with an Intel(R) Xeon(R) CPU E5-2690 v4, That is, drec→GT measures a notion of precision of the re-
256 GBgb of memory, and 4 NVIDIA Tesla P40 GPUs. The construction, while dinp→rec measures a notion of recall.
runtime of our algorithm are considerably higher than com- Whereas drec→GT is an indication of overall quality, it does
peting methods, requiring around 0.45 minutes per patch, not penalize the methods for not covering undersampled re-
for a total of up to 6.5 hours to optimize an the entire model. gions of the input. Figure 6 illustrates these one-sided corre-
We optimize the losses (3) and (4) using the ADAM spondence measures. A naive reconstruction that copies the
[23] implementation in PyTorch with default parameters. input satisfies dinp→rec ≡ 0 but since in general the input
Specifically for ADAM, we use a learning rate of 10−3 , point cloud consists of noisy measurements, we will have
β = (0.9, 0.999),  = 10−8 , and no weight decay. For the dinp→GT > 0.
 2%

1.5%

1%

0.5%

0%
Anchor Dancing Children Daratech Gargoyle Lord Quas

Figure 4. Fitted points on the different models of the benchmark, the color illustrates the error with respect to the ground truth.

1.0e+0 Apss 1.0e+0 Apss

Ear Ear
Fourier Fourier
8.0e−1 Imls 8.0e−1 Imls
Mpu Mpu
Mpusmooth Mpusmooth
Poisson Poisson
6.0e−1 6.0e−1

Frequency

Frequency
Screen Poisson Screen Poisson
Rbf Rbf
Scattered Scattered
4.0e−1 Spss 4.0e−1 Spss
Wavelet Wavelet
Our Our
2.0e−1 2.0e−1

0.0e+0 0.0e+0

0 0.001 0.002 0.003 0.004 0.005 0 0.001 0.002 0.003 0.004 0.005

Error Error

Anchor Gargoyle

First patches Final patches Figure 8. Percentage of fitted vertices (y-axis) to reach a given er-
Figure 5. Initial patch fitting (left), we clearly see the “un-mixed” ror (drec→GT ), x-axis) for different methods. The errors are mea-
colors due to the disagreeing patches. After optimization (right) sured as distance from the input data to the fitted surface. The
all patches are overlapping producing a mix of color. plots for the remaining models of the dataset are provided in the
supplementary document.

dinp!rec drec!GT

Input Reconstructed Ground Truth

Figure 6. Diagram illustrating the evaluation measures we use to

compare different reconstructions.

1.0e+0 Apss 1.0e+0 Apss

1.0e−1
Ear
Fourier
1.0e−1
Ear
Fourier
EAR Ours
Imls Imls
Mpu Mpu
1.0e−2 1.0e−2
Mpusmooth Mpusmooth
Poisson Poisson
Figure 9. EAR (left) versus Ours (right). The input point cloud
Frequency

Frequency

1.0e−3 Screen Poisson 1.0e−3 Screen Poisson

Rbf Rbf
1.0e−4 Scattered
Spss
1.0e−4 Scattered
Spss
contains noise which is smoothed out by our reconstruction (while
1.0e−5 Wavelet Wavelet
1.0e−5

1.0e−6
Our

1.0e−6
Our
preserving sharp features quite well) but interpolated by EAR. The
1.0e−7 1.0e−7 result is that EAR produces spurious points and visible artifacts.
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1

Error Error

Anchor Gargoyle

Figure 7. Percentage of fitted vertices (y-axis, log scale) to reach

and it is on par with the state-of-the art EAR method [19],
a given error (dinp→rec , x-axis) for different methods. The errors which achieves a similar score for these 5 models. But, as
are computed from the fitted surface to the ground truth. The plots we will discuss in the next paragraph, EAR is unable to cope
for the remaining models of the dataset are provided in the supple- with noisy inputs. This is a remarkable result considering
mentary document. that our method, differently from all the others, produces an
explicit parametrization of the surface, which can be resam-
pled at arbitrary resolutions, used for texture mapping, or to
Figures 7 and 8 show respectively the percentage of ver- compute analytic normals and curvature values.
tices of Yb and X such that drec→GT and dinp→rec is below EAR is an interpolative method, which by construction
a given error. satisfies X ⊂ Y. b It follows that the noise in the measure-
Our technique outperforms all the technique we tested, ments is necessarily transferred to the surface reconstruc-
 Sharp Features

Noise Chamfer

Reference Poisson Ear Our

Figure 10. Example of reconstruction on extreme conditions for Wasserstein (w/ Sinkhorn)
our method versus the most representative traditional methods.
Figure 11. Surface reconstruction of stacked cubes using a sin-
The red dots in the Reference are the input points for the recon-
gle chart, with two different choices of metric. We verify
struction.
that the Wasserstein metric, even with the Sinkhorn entropy-
regularization, provides a more uniform parametrisation, as well
as bijective correspondences between overlapping charts.
tion. Figure 9 illustrates that our deep geometric prior pre-
serves the sharp structures while successfully denoising the
input point cloud. The mathematical analysis of such im-
Net is a data-driven reconstruction method using an autoen-
plicit regularisation is a fascinating direction of future work.
coder architecture. While their emphasis was on leverag-
ing semantic similarity of shapes and images on several 3D
Noise and Sharp Features. As discussed above, the be- tasks, we focus on high-fidelity point cloud reconstruction
havior of surface reconstruction methods is particularly in data-sparse regimes, i.e. in absence of any training data.
challenging in the presence of sharp features and noise in Our main contribution is to show that in such regimes, an
the input measurements. We performed two additional ex- even simpler neural network yields state-of-the-art results
periments to compare the behaviour of our architecture with on surface reconstruction. We also note the following es-
the most representative traditional methods on both noisy sential differences between our method and AtlasNet.
point clouds and models with sharp features. Schreened
Poisson Surface Reconstruction [22] (Figure 10, left) is very • No Learning: Our model does not require any training
robust to noise, but fails at capturing sharp features. EAR data, and, as a result, we do not need to consider an
(Figure 10, middle) is the opposite: it captures sharp fea- autoencoder architecture with specific encoders.
tures accurately, but being an interpolatory method fails
at reconstructing noisy inputs, thus introducing spurious • Transition Functions: Since we have pointwise cor-
points and regions during the reconstruction. Our method respondences, we can define a transition function be-
(Figure 10, right) does not suffer from these limitations, ro- tween overlapping patches Vp and Vq by consistently
bustly fitting noisy inputs and capturing sharp features. triangulating corresponding parametric points in Vp
and Vq . In contrast, AtlasNet does not have such corre-
Generating a Triangle Mesh. Our method generates spondences and thus does not produce a true manifold
a collection of local charts, which can be sampled atlas.
at an arbitrary resolution. We can generate a trian- • Patch Selection: We partition the input into point-sets
gle mesh by using off-the-shelf-techniques such as Pois- Xp that we fit separately. While it is theoretically at-
son Surface Reconstruction [22] on our dense point tractive to attempt to fit each patch to the whole set as
clouds. We provide meshes reconstructed in this way it is done in AtlasNet, and let the algorithm figure out
for all the benchmark models at https://github.com/ the patch partition automatically, in practice the diffi-
fwilliams/deep-geometric-prior. culty of the optimization problem leads to unsatisfac-
tory results. In other words, AtlasNet is approximating
Comparison with AtlasNet [12]. Our atlas construction a global matching whereas our model only requires lo-
is related to the AtlasNet model introduced in [12]. Atlas- cal matchings within each patch.
 Atlasnet Atlasnet
1.0e+0
5.0e−1
2.0e−1
1.0e−1
Our
1.0e+0

8.0e−1
Our
5. Discussion
5.0e−2
2.0e−2
1.0e−2 6.0e−1
Neural networks – particularly in the overparametrised
Frequency

Frequency
5.0e−3
2.0e−3
1.0e−3
5.0e−4
2.0e−4
4.0e−1
regime – are remarkably efficient at curve fitting in high-
1.0e−4
5.0e−5
2.0e−5
2.0e−1
dimensional spaces. Despite recent progress in understand-
1.0e−5

0 0.02 0.04 0.06 0.08 0.1

0.0e+0

0 0.001 0.002 0.003 0.004 0.005 ing the dynamics of gradient descent in such regimes, their
Error Error

ability to learn and generalize by overcoming the curse

of dimensionality remains a major mystery. In this pa-
per, we bypass this high-dimensional regime and concen-
trate on a standard low-dimensional interpolation problem:
surface reconstruction. We demonstrate that in this regime
neural networks also have an intriguing capacity to recon-
struct complex signal structures while providing robustness
to noise.
Our model is remarkably simple, combining two key
principles: (i) construct local piece-wise linear charts by
Atlasnet Our
means of a vanilla ReLU fully-connected network, and (ii)
Figure 12. Comparison with Atlasnet. Top: drec→GT and dinp→rec use Wasserstein distances in each neighborhood, enabling
cumulative histograms. Bottom: surface reconstruction visualiza- consistent transitions across local charts. The resulting ar-
tion. Observe that local charts do not have consistent transitions in chitecture, when combined with gradient descent, provides
the Atlasnet output. a “deep geometric prior” that is shown to outperform exist-
ing surface-reconstruction methods, which rely on domain-
specific geometric assumptions. The theoretical analysis of
• Wasserstein vs. Chamfer Distance: As discussed this deep geometric prior is our next focus, which should
above, the EMD automatically provides transition address questions such as how the geometry of the surface
maps across local charts. AtlasNet considers instead informs the design the neural network architecture, or why
the Chamfer distance between point clouds, which is gradient descent biasing towards locally regular recon-
is more efficient to compute but sacrifices the abil- structions.
ity to construct bijections in the overlapping regions. Despite these promising directions, we also note the lim-
Moreover, as illustrated in Figure 11, we observe that itations our approach is facing. In particular, our method
Chamfer distances may result in distortion effects even is currently substantially more expensive than alternatives.
within local charts. One natural possibility to accelerate it, would be to train a
separate neural network model to provide an efficient ini-
tialization for the local chart minimisation (2), similarly as
• Chart Consistency: We explicitly enforce consis-
in neural sparse coders [11]. Another important question for
tency (4) which has a significant effect on quality, as
future research is the adaptive patch selection, which would
illustrated in Section 4, whereas AtlasNet does not pro-
leverage the benefits of multiscale approximations.
duce a real manifold atlas, since it has no definition of
transition maps.
6. Acknowledgements
We are grateful to the NYU HPC staff for providing
We provide quantitative and qualitative comparisons to computing cluster service. This project was partially sup-
assess the impact of our architecture choices by adapting ported by the NSF CAREER award 1652515, the NSF
AtlasNet to a data-free setting. In this setting, we overfit At- grant IIS-1320635, the NSF grant DMS-1436591, the NSF
lasNet on a single model with the same number of patches grant DMS-1835712, the NSF grant RI-IIS 1816753, the
used for our method. Figure 12 reports both drec→GT and SNSF grant P2TIP2 175859, the Alfred P. Sloan Foun-
dinp→rec cumulative histograms on a twisted cube surface dation, the Moore-Sloan Data Science Environment, the
using 10 local charts. We verify that when the Atlasnet DARPA D3M program, NVIDIA, Samsung Electronics,
architecture is trained to fit the surface using our experi- Labex DigiCosme, DOA W911NF-17-1-0438, a gift from
mental setup, it is clearly outperformed both quantitatively Adobe Research, and a gift from nTopology. Any opinions,
and qualitatively by our deep geometric prior. We empha- findings, and conclusions or recommendations expressed in
size however that AtlasNet is designed as a data-driven ap- this material are those of the authors and do not necessarily
proach, and as such it can leverage semantic information reflect the views of DARPA. The authors would also like to
from large training sets. thank the anonymous reviewers for their time and effort.
References point clouds. Computer Graphics Forum, 37(2):75–
85, 2018. 2
[1] A. Adamson and M. Alexa. Approximating and inter-
secting surfaces from points. In SGP, page 239, 2003. [15] S. Gunasekar, J. Lee, D. Soudry, and N. Srebro. Char-
2 acterizing implicit bias in terms of optimization geom-
etry. arXiv preprint arXiv:1802.08246, 2018. 4
[2] M. Alexa, J. Behr, D. Fleishman, D. Levin, and
C. Silva. Point Set Surfaces. In IEEE VIS, page 21, [16] X. Han, Z. Li, H. Huang, E. Kalogerakis, and Y. Yu.
2001. 2 High-resolution shape completion using deep neural
networks for global structure and local geometry in-
[3] J. Altschuler, J. Weed, and P. Rigollet. Near-linear ference. In Proceedings of IEEE International Con-
time approximation algorithms for optimal transport ference on Computer Vision (ICCV), 2017. 2
via sinkhorn iteration. In Advances in Neural Infor-
mation Processing Systems, pages 1964–1974, 2017. [17] H. Hoppe, T. DeRose, T. Duchamp, J. McDonald, and
3 W. Stuetzle. Surface reconstruction from unorganized
points. In SIGGRAPH, pages 71–78, 1992. 2
[4] R. Basri and D. Jacobs. Efficient representation
of low-dimensional manifolds using deep networks. [18] H. Huang, S. Wu, M. Gong, D. Cohen-Or, U. M. As-
arXiv preprint arXiv:1602.04723, 2016. 2 cher, and H. R. Zhang. Edge-aware point set resam-
pling. ACM TOG, 32(1):9:1–9:12, 2013. 2
[5] M. Berger, J. A. Levine, L. G. Nonato, G. Taubin, and
C. T. Silva. A benchmark for surface reconstruction. [19] Q. Huang, F. Wang, and L. Guibas. Functional map
ACM TOG, 32(2):20:1–20:17, 2013. 2, 5, 11 networks for analyzing and exploring large shape col-
lections. In SIGGRAPH, 2014. 6
[6] M. Berger, A. Tagliasacchi, L. M. Seversky, P. Al-
liez, G. Guennebaud, J. A. Levine, A. Sharf, and C. T. [20] M. Kazhdan. Reconstruction of solid models from ori-
Silva. A survey of surface reconstruction from point ented point sets. In SGP, page 73, 2005. 2
clouds. CGF, 36(1):301–329, 2017. 2 [21] M. Kazhdan, M. Bolitho, and H. Hoppe. Poisson sur-
[7] J. Bowers, R. Wang, L.-Y. Wei, and D. Maletz. Par- face reconstruction. In SGP, page 70, 2006. 2
allel poisson disk sampling with spectrum analysis on [22] M. M. Kazhdan and H. Hoppe. Screened poisson
surfaces. In ACM Transactions on Graphics (TOG), surface reconstruction. ACM TOG, 32(3):29:1–29:13,
volume 29, page 166. ACM, 2010. 4 2013. 2, 7
[8] M. Cuturi. Sinkhorn distances: Lightspeed computa- [23] D. P. Kingma and J. Ba. Adam: A method for stochas-
tion of optimal transport. In Advances in neural infor- tic optimization. arXiv preprint arXiv:1412.6980,
mation processing systems, pages 2292–2300, 2013. 2014. 5
3 [24] R. Kolluri. Provably Good Moving Least Squares. In
[9] A. Dai, C. R. Qi, and M. Nießner. Shape completion SODA, pages 1008–1018, 2005. 2
using 3d-encoder-predictor cnns and shape synthesis. [25] H. W. Kuhn. The hungarian method for the assign-
In Proc. IEEE Conf. on Computer Vision and Pattern ment problem. Naval research logistics quarterly, 2(1-
Recognition (CVPR), volume 3, 2017. 2 2):83–97, 1955. 3
[10] S. S. Du, X. Zhai, B. Poczos, and A. Singh. Gradient [26] Y. Li, T. Ma, and H. Zhang. Algorithmic regular-
descent provably optimizes over-parameterized neural ization in over-parameterized matrix recovery. arXiv
networks. arXiv preprint arXiv:1810.02054, 2018. 3 preprint arXiv:1712.09203, 2017. 4
[11] K. Gregor and Y. LeCun. Learning fast approxima- [27] J. Manson, G. Petrova, and S. Schaefer. Streaming sur-
tions of sparse coding. In Proceedings of the 27th face reconstruction using wavelets. 27(5):1411–1420,
International Conference on International Conference 2008. 2
on Machine Learning, pages 399–406. Omnipress, [28] Y. Nagai, Y. Ohtake, and H. Suzuki. Smoothing of
2010. 8 Partition of Unity Implicit Surfaces for Noise Robust
[12] T. Groueix, M. Fisher, V. G. Kim, B. C. Russell, and Surface Reconstruction. 28(5):1339–1348, 2009. 2
M. Aubry. Atlasnet: A papier-m\ˆ ach\’e approach [29] Y. Ohtake, A. Belyaev, M. Alexa, G. Turk, and H. Sei-
to learning 3d surface generation. arXiv preprint del. Multi-level partition of unity implicits. ACM
arXiv:1802.05384, 2018. 2, 7 TOG, 22(3):463–470, 2003. 2
[13] G. Guennebaud and M. Gross. Algebraic Point Set [30] Y. Ohtake, A. Belyaev, and H. Seidel. An integrating
Surfaces. SIGGRAPH, 26:23–9, 2007. 2 approach to meshing scattered point data. In Sympo-
[14] P. Guerrero, Y. Kleiman, M. Ovsjanikov, and N. J. Mi- sium on Solid and Physical Modeling, page 69, 2005.
tra. PCPNet: Learning local shape properties from raw 2
[31] Y. Ohtake, A. G. Belyaev, and H.-P. Seidel. 3d scat-
tered data interpolation and approximation with mul-
tilevel compactly supported rbfs. Graphical Models,
67(3):150–165, 2005. 2
[32] N. Pietroni, M. Tarini, O. Sorkine, and D. Zorin.
Global parametrization of range image sets. In ACM
Transactions on Graphics (TOG), volume 30, page
149. ACM, 2011. 2
[33] C. R. Qi, H. Su, K. Mo, and L. J. Guibas. Pointnet:
Deep learning on point sets for 3d classification and
segmentation. In CVPR, 2017. 2
[34] C. R. Qi, L. Yi, H. Su, and L. J. Guibas. Pointnet++:
Deep hierarchical feature learning on point sets in a
metric space. arXiv preprint arXiv:1706.02413, 2017.
2
[35] D. Soudry, E. Hoffer, M. S. Nacson, S. Gunasekar,
and N. Srebro. The implicit bias of gradient descent
on separable data. arXiv preprint arXiv:1710.10345,
2017. 4
[36] S. Sukhbaatar, R. Fergus, et al. Learning multiagent
communication with backpropagation. In Advances in
Neural Information Processing Systems, pages 2244–
2252, 2016. 2
[37] D. Ulyanov, A. Vedaldi, and V. Lempitsky. Deep im-
age prior. arXiv preprint arXiv:1711.10925, 2017. 1,
3
[38] J. Varley, C. DeChant, A. Richardson, J. Ruales, and
P. Allen. Shape completion enabled robotic grasp-
ing. In Intelligent Robots and Systems (IROS), 2017
IEEE/RSJ International Conference on, pages 2442–
2447. IEEE, 2017. 2
[39] O. Vinyals, S. Bengio, and M. Kudlur. Order mat-
ters: Sequence to sequence for sets. arXiv preprint
arXiv:1511.06391, 2015. 2
[40] S. Xiong, J. Zhang, J. Zheng, J. Cai, and L. Liu.
Robust surface reconstruction via dictionary learning.
ACM Transactions on Graphics (TOG), 33(6):201,
2014. 2
[41] L. Yu, X. Li, C.-W. Fu, D. Cohen-Or, and P.-A. Heng.
Pu-net: Point cloud upsampling network. In Proceed-
ings of the IEEE Conference on Computer Vision and
Pattern Recognition, pages 2790–2799, 2018. 2
 λ = 0.001 λ = 0.005 λ = 0.01 λ = 0.05 λ = 0.1 λ = 0.5 λ=1
W2 = 2.46 W2 = 2.87 W2 = 3.26 W2 = 10.76 W2 = 20.98 W2 = 92.91 W2 = 174.09

Figure 13. Effect of the parameter λ on the reconstruction and the loss W2 .

A.3. Surface Reconstruction Benchmark

We provide cumulative histograms for the results of the
Surface Reconstruction Benchmark [5] on all 5 models
shown in Figure 4. Figures 15 and 16 show respectively
the percentage of vertices of Yb and X such that drec→GT
and dinp→rec is below a given error.
A.4. Surface Reconstruction Benchmark Statistics
In addition to the cumulative histograms above, we tabu-
late the mean, standard deviation, and maximum values for
each method and model in the benchmark. Table 1 show
the distance from the input to the reconstruction (dinp→rec )
and Table 2 show the distance from the reconstruction to the
input (drec→GT ).
Figure 14. Example of reconstruction form Kinect data (red dots).

A. Supplementary Experiments
A.1. Effect of the parameter λ
In Figure 13, we demonstrate the effect of varying the
Sinkhorn regularization parameter on the final reconstruc-
tion of a surface. Smaller values of λ yield a better approxi-
mation of the Wasserstein distance, and thus, produce better
reconstructions of the original points.
A.2. Kinect reconstruction
To demonstrate the effectiveness of our technique on re-
constructing point clouds with large quantities of noise and
highly non-uniform sampling, we reconstruct a raw point
cloud acquired with a Kinect V2 (Figure 14). In spite of
the challenging input, we are still able to produce a smooth
reconstruction approximating the geometry of the original
object.
 1.0e+0 Apss 1.0e+0 Apss
Ear Ear
Fourier Fourier
1.0e−1 1.0e−1
Imls Imls
Mpu Mpu
1.0e−2 1.0e−2
Mpusmooth Mpusmooth
Poisson Poisson
Frequency

Frequency
1.0e−3 Screen Poisson 1.0e−3 Screen Poisson
Rbf Rbf
1.0e−4 Scattered 1.0e−4 Scattered
Spss Spss
1.0e−5 Wavelet Wavelet
1.0e−5
Our Our

1.0e−6 1.0e−6

1.0e−7 1.0e−7

0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1

Error Error

Anchor Gargoyle

1.0e+0 Apss 1.0e+0 Apss 1.0e+0 Apss

Ear Ear Ear
Fourier Fourier Fourier
1.0e−1 1.0e−1 1.0e−1
Imls Imls Imls
Mpu Mpu Mpu
1.0e−2 1.0e−2 1.0e−2
Mpusmooth Mpusmooth Mpusmooth
Poisson Poisson Poisson
Frequency

Frequency

Frequency
1.0e−3 Screen Poisson 1.0e−3 Screen Poisson 1.0e−3 Screen Poisson
Rbf Rbf Rbf
Scattered 1.0e−4 Scattered 1.0e−4 Scattered
1.0e−4
Spss Spss Spss
Wavelet 1.0e−5 Wavelet Wavelet
1.0e−5 1.0e−5
Our Our Our

1.0e−6 1.0e−6
1.0e−6

1.0e−7 1.0e−7
1.0e−7
0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1 0 0.02 0.04 0.06 0.08 0.1

Error Error Error

Daratech Danching Children Lord Quas

Figure 15. Percentage of fitted vertices (y-axis, log scale) to reach a certain error level (x-axis) for different methods. The errors are
computed from the fitted surface to the ground truth.

1.0e+0 Apss 1.0e+0 Apss

Ear Ear
Fourier Fourier
8.0e−1 Imls 8.0e−1 Imls
Mpu Mpu
Mpusmooth Mpusmooth
Poisson Poisson
6.0e−1 6.0e−1
Frequency

Frequency

Screen Poisson Screen Poisson

Rbf Rbf
Scattered Scattered
4.0e−1 Spss 4.0e−1 Spss
Wavelet Wavelet
Our Our
2.0e−1 2.0e−1

0.0e+0 0.0e+0

0 0.001 0.002 0.003 0.004 0.005 0 0.001 0.002 0.003 0.004 0.005

Error Error

Anchor Gargoyle

1.0e+0 Apss 1.0e+0 Apss 1.0e+0 Apss

Ear Ear Ear
Fourier Fourier Fourier
8.0e−1 Imls 8.0e−1 Imls 8.0e−1 Imls
Mpu Mpu Mpu
Mpusmooth Mpusmooth Mpusmooth
Poisson Poisson Poisson
6.0e−1 6.0e−1 6.0e−1
Frequency

Frequency

Frequency

Screen Poisson Screen Poisson Screen Poisson

Rbf Rbf Rbf
Scattered Scattered Scattered
4.0e−1 Spss 4.0e−1 Spss 4.0e−1 Spss
Wavelet Wavelet Wavelet
Our Our Our
2.0e−1 2.0e−1 2.0e−1

0.0e+0 0.0e+0 0.0e+0

0 0.001 0.002 0.003 0.004 0.005 0 0.001 0.002 0.003 0.004 0.005 0 0.001 0.002 0.003 0.004 0.005

Error Error Error

Daratech Danching Children Lord Quas

Figure 16. Percentage of fitted vertices (y-axis) to reach a certain error level (x-axis) for different methods. The errors are measured as
distance from the input data to the fitted surface.
 min avg std max min avg std max
Apss 1.32e-05 8.78e-04 8.78e-04 9.21e-03 Apss 6.28e-06 1.79e-03 1.79e-03 2.80e-02
Ear 3.30e-09 3.18e-07 3.18e-07 7.75e-07 Ear 4.96e-06 1.53e-03 1.53e-03 9.93e-03
Fourier 8.95e-06 1.34e-03 1.34e-03 1.34e-01 Fourier 1.45e-06 2.01e-03 2.01e-03 6.59e-02
Imls 1.11e-05 9.00e-04 9.00e-04 9.43e-03 Imls 6.49e-06 1.92e-03 1.92e-03 2.82e-02
Mpu 4.07e-06 8.73e-04 8.73e-04 8.45e-03 Mpu 2.23e-06 2.08e-03 2.08e-03 4.59e-02
Anchor

Anchor
Mpusmooth 9.50e-06 9.13e-04 9.13e-04 1.06e-02 Mpusmooth 4.81e-06 1.87e-03 1.87e-03 3.66e-02
Poisson 1.09e-05 1.63e-03 1.63e-03 1.17e-01 Poisson 8.75e-06 2.27e-03 2.27e-03 6.59e-02
Screen Poisson 6.99e-06 7.45e-04 7.45e-04 1.79e-02 Screen Poisson 7.12e-06 2.15e-03 2.15e-03 6.59e-02
Rbf 1.46e-05 8.61e-04 8.61e-04 9.89e-03 Rbf 1.62e-06 2.98e-03 2.98e-03 6.60e-02
Scattered 1.33e-05 8.18e-04 8.18e-04 1.06e-02 Scattered 4.30e-06 2.16e-03 2.16e-03 6.57e-02
Spss 8.67e-06 1.03e-03 1.03e-03 1.06e-02 Spss 4.15e-06 4.39e-03 4.39e-03 9.00e-02
Wavelet 8.30e-06 2.19e-03 2.19e-03 6.27e-02 Wavelet 5.36e-06 3.01e-03 3.01e-03 6.59e-02
Our 4.91e-06 7.21e-04 7.21e-04 2.55e-02 Our 3.82e-06 1.53e-03 1.53e-03 9.69e-03
Apss 9.98e-06 7.87e-04 7.87e-04 1.05e-02 Apss 2.18e-06 1.68e-03 1.68e-03 2.16e-02
Ear 1.34e-09 2.79e-07 2.79e-07 8.18e-07 Ear 3.15e-06 1.50e-03 1.50e-03 1.68e-02
Fourier 7.86e-06 1.06e-03 1.06e-03 1.94e-02 Fourier 4.68e-06 2.39e-03 2.39e-03 5.98e-02
Imls 5.73e-06 8.35e-04 8.35e-04 1.05e-02 Imls 3.46e-06 1.79e-03 1.79e-03 2.36e-02
Mpu 5.33e-06 8.47e-04 8.47e-04 8.55e-03 Mpu 5.25e-06 2.06e-03 2.06e-03 4.11e-02
Daratech

Daratech
Mpusmooth 9.87e-06 9.31e-04 9.31e-04 1.87e-02 Mpusmooth 5.98e-06 2.14e-03 2.14e-03 4.92e-02
Poisson 1.28e-05 1.58e-03 1.58e-03 3.18e-02 Poisson 3.25e-06 2.98e-03 2.98e-03 6.83e-02
Screen Poisson 3.80e-06 6.98e-04 6.98e-04 1.72e-02 Screen Poisson 4.30e-06 2.42e-03 2.42e-03 6.47e-02
Rbf 2.12e-06 7.52e-04 7.52e-04 1.08e-02 Rbf 2.28e-06 3.59e-03 3.59e-03 6.44e-02
Scattered 7.48e-06 6.97e-04 6.97e-04 1.70e-02 Scattered 5.40e-06 1.60e-03 1.60e-03 1.81e-02
Spss 8.36e-06 1.12e-03 1.12e-03 1.12e-02 Spss 5.99e-06 2.73e-03 2.73e-03 7.66e-02
Wavelet 6.13e-06 1.88e-03 1.88e-03 2.27e-02 Wavelet 4.04e-06 3.29e-03 3.29e-03 5.04e-02
Our 5.30e-06 4.23e-04 4.23e-04 1.79e-02 Our 1.96e-06 1.51e-03 1.51e-03 1.67e-02
Apss 6.20e-06 4.98e-04 4.98e-04 1.45e-02 Apss 5.53e-06 1.68e-03 1.68e-03 2.87e-02
Ear 9.81e-10 3.18e-07 3.18e-07 8.28e-07 Ear 2.54e-06 1.32e-03 1.32e-03 2.46e-02
Fourier 7.69e-06 6.23e-04 6.23e-04 2.65e-02 Fourier 3.51e-06 1.60e-03 1.60e-03 2.53e-02
Imls 9.09e-06 5.88e-04 5.88e-04 1.43e-02 Imls 6.12e-06 1.75e-03 1.75e-03 2.93e-02
Mpu 1.07e-05 5.54e-04 5.54e-04 7.08e-03 Mpu 3.98e-06 1.53e-03 1.53e-03 3.84e-02
Mpusmooth 8.10e-06 6.14e-04 6.14e-04 2.75e-02 Mpusmooth 3.51e-06 1.47e-03 1.47e-03 1.64e-02
Dc

Dc

Poisson 6.76e-06 1.02e-03 1.02e-03 2.63e-02 Poisson 5.86e-06 1.87e-03 1.87e-03 2.53e-02
Screen Poisson 7.12e-06 4.34e-04 4.34e-04 2.70e-02 Screen Poisson 3.96e-06 1.49e-03 1.49e-03 2.25e-02
Rbf 9.15e-06 6.40e-04 6.40e-04 2.77e-02 Rbf 1.15e-06 1.55e-03 1.55e-03 3.01e-02
Scattered 4.45e-06 5.20e-04 5.20e-04 2.69e-02 Scattered 5.26e-06 1.89e-03 1.89e-03 5.27e-02
Spss 3.76e-06 7.66e-04 7.66e-04 1.55e-02 Spss 5.12e-06 1.96e-03 1.96e-03 2.93e-02
Wavelet 1.76e-05 1.82e-03 1.82e-03 2.68e-02 Wavelet 4.70e-06 2.63e-03 2.63e-03 2.56e-02
Our 6.10e-06 3.98e-04 3.98e-04 2.48e-02 Our 3.10e-06 1.31e-03 1.31e-03 1.51e-02
Apss 7.80e-06 5.62e-04 5.62e-04 6.92e-03 Apss 2.64e-06 1.50e-03 1.50e-03 3.41e-02
Ear 1.73e-09 3.18e-07 3.18e-07 7.52e-07 Ear 4.04e-06 1.22e-03 1.22e-03 8.81e-03
Fourier 3.94e-06 7.02e-04 7.02e-04 2.55e-02 Fourier 2.29e-06 1.37e-03 1.37e-03 2.15e-02
Imls 8.39e-06 6.09e-04 6.09e-04 6.49e-03 Imls 2.11e-06 1.71e-03 1.71e-03 4.37e-02
Mpu 1.10e-05 6.27e-04 6.27e-04 6.75e-03 Mpu 4.60e-06 1.57e-03 1.57e-03 2.98e-02
Gargoyle

Gargoyle

Mpusmooth 4.57e-06 7.25e-04 7.25e-04 8.81e-03 Mpusmooth 1.20e-06 1.37e-03 1.37e-03 2.17e-02
Poisson 1.20e-05 1.05e-03 1.05e-03 2.73e-02 Poisson 6.48e-06 1.57e-03 1.57e-03 2.17e-02
Screen Poisson 1.04e-05 4.87e-04 4.87e-04 1.81e-02 Screen Poisson 2.82e-06 1.30e-03 1.30e-03 2.09e-02
Rbf 6.44e-06 7.30e-04 7.30e-04 5.86e-03 Rbf 3.19e-06 7.66e-03 7.66e-03 9.17e-02
Scattered 7.73e-06 5.78e-04 5.78e-04 1.03e-02 Scattered 2.48e-06 1.36e-03 1.36e-03 2.17e-02
Spss 5.30e-06 7.74e-04 7.74e-04 1.34e-02 Spss 5.03e-06 1.85e-03 1.85e-03 4.65e-02
Wavelet 1.14e-05 1.56e-03 1.56e-03 2.73e-02 Wavelet 4.79e-06 1.89e-03 1.89e-03 2.11e-02
Our 5.40e-06 4.50e-04 4.50e-04 1.81e-02 Our 2.34e-06 1.19e-03 1.19e-03 1.45e-02
Apss 8.64e-06 4.76e-04 4.76e-04 7.55e-03 Apss 3.90e-06 1.24e-03 1.24e-03 2.14e-02
Ear 1.05e-09 3.22e-07 3.22e-07 8.91e-07 Ear 2.26e-06 1.14e-03 1.14e-03 7.49e-03
Fourier 1.15e-05 5.64e-04 5.64e-04 1.79e-02 Fourier 5.01e-06 1.30e-03 1.30e-03 1.96e-02
Imls 8.29e-06 5.29e-04 5.29e-04 8.60e-03 Imls 3.30e-06 1.31e-03 1.31e-03 2.36e-02
Mpu 7.94e-06 5.44e-04 5.44e-04 5.01e-03 Mpu 2.30e-06 1.35e-03 1.35e-03 2.94e-02
Lord Quas

Lord Quas

Mpusmooth 1.07e-05 5.70e-04 5.70e-04 8.18e-03 Mpusmooth 4.70e-06 1.28e-03 1.28e-03 1.54e-02
Poisson 5.70e-06 8.24e-04 8.24e-04 4.38e-02 Poisson 1.91e-06 1.39e-03 1.39e-03 1.94e-02
Screen Poisson 4.74e-06 4.29e-04 4.29e-04 1.08e-02 Screen Poisson 2.07e-06 1.24e-03 1.24e-03 1.62e-02
Rbf 1.01e-05 6.48e-04 6.48e-04 7.27e-03 Rbf 2.26e-06 1.29e-03 1.29e-03 1.80e-02
Scattered 6.72e-06 4.88e-04 4.88e-04 1.69e-02 Scattered 4.48e-06 1.17e-03 1.17e-03 1.32e-02
Spss 6.66e-06 6.03e-04 6.03e-04 7.66e-03 Spss 1.49e-06 1.38e-03 1.38e-03 2.28e-02
Wavelet 9.27e-06 1.49e-03 1.49e-03 4.71e-02 Wavelet 4.42e-06 1.87e-03 1.87e-03 1.52e-02
Our 1.38e-06 4.15e-04 4.15e-04 2.14e-02 Our 3.94e-06 1.14e-03 1.14e-03 8.74e-03

Table 1. Distance from the input to the reconstruction. Table 2. Distance from the reconstruction to the input.