Thingi10K: A Dataset of 10,000 3D-Printing Models

Qingnan Zhou Alec Jacobson

New York University Columbia University
arXiv:1605.04797v2 [cs.GR] 2 Jul 2016

Figure 1: The Thingi10K dataset contains 10,000 models from from featured “things” on thingiverse.com, a popular online repository.

Abstract the demand for state-of-the-art processing techniques and automa-

tion within 3D printing pipelines.
Empirically validating new 3D-printing related algorithms and im-
plementations requires testing data representative of inputs encoun- However, testing remains inadequate. Existing datasets contain
tered in the wild. An ideal benchmarking dataset should not only only sanitized models (e.g., [Aim@Shape 2004; Levoy et al. 2005;
draw from the same distribution of shapes people print in terms of Myles et al. 2014]) or draw from populations containing raw mod-
class (e.g., toys, mechanisms, jewelry), representation type (e.g., els not specifically intended for printing (rather, e.g., for shape clas-
triangle soup meshes) and complexity (e.g., number of facets), but sification [Shilane et al. 2004; Chang et al. 2015] or scene under-
should also capture problems and artifacts endemic to 3D print- standing [Nathan Silberman & Fergus 2012; Choi et al. 2016]).
ing models (e.g., self-intersections, non-manifoldness). We observe In this paper, we will show that the characteristics and issues com-
that the contextual and geometric characteristics of 3D printing mon to 3D printing models are distinct from models intended for
models differ significantly from those used for computer graphics visualization. As such, validating geometry processing techniques
applications, not to mention standard models (e.g., Stanford bunny, related to 3D printing requires a new representative dataset. This
Armadillo, Fertility). We present a new dataset of 10,000 models ideal dataset should encompass the different contextual and geo-
collected from an online 3D printing model-sharing database. Via metric characteristics of commonly printed shapes. Characteristics
analysis of both geometric (e.g., triangle aspect ratios, manifold- common to 3D printing models should appear with proportional
ness) and contextual (e.g., licenses, tags, classes) characteristics, distributions, and characteristics inconsistent with models intended
we demonstrate that this dataset represents a more concise sum- for fabrication should be infrequent (e.g., the open boundaries of a
mary of real-world models used for 3D printing compared to exist- video game character’s clothing).
ing datasets. To facilitate future research endeavors, we also present
an online query interface to select subsets of the dataset according We propose a dataset of 10,000 models culled from a popular shape
to project-specific characteristics. The complete dataset and per- repository for 3D printing enthusiasts, thingiverse.com. Hereon,
model statistical data are freely available to the public. we refer to our dataset as Thingi10K. Beyond collecting tags and
class information available online, we analyze geometric character-
istics of each model (e.g., manifoldness, lack of self-intersections,
1 Introduction and background genus). We contrast these statistics against existing large datasets
and investigate correlations within the data.
The iconic Stanford bunny, now 23 years old, has been melted,
shattered, and deformed countless times. While mostly a fun sub- Existing datasets. Myles et al. collect 116 models from aca-
culture, “bunny torture” is also a legacy of an earlier time when demic sources (Stanford Scanning Repository [Levoy et al. 2005]
few interesting and free 3D models existed. Testing on such stan- and Aim@Shape Repository [Aim@Shape 2004]) to test their pa-
dard models persists despite well-known limitations. As Greg Turk, rameterization algorithm [Myles et al. 2014]. These models corre-
originator of the bunny, advises, “I actually consider the bunny to spond to best-case input due to their extreme cleanliness and gen-
be too good as a test model. It is fairly smooth, it has manifold eral position assumption (i.e., no four points on a circle, no coplanar
connectivity, and it isn’t too complex” [Turk 2000]. intersections, etc.). For 3D printing models in the wild, degenera-
cies, non-manifoldness and self-intersections are abundant, not spe-
Oversimplified testing provides a false sense of robustness and cial cases. Structured modeling and coordinate quantization tends
causes not only visual artifacts in computer graphics applications, to break rather than fulfill general position assumptions.
but also fabrication and functionality artifacts when processing ge-
ometry intended for 3D printing. Fortunately, 3D models are now Computer vision and machine learning applications demand large
abundant. Modern consumer-level 3D printing technologies nurture scale training datasets. For example, the NYU Depth Dataset col-
new communities of professional and amateur 3D modelers, who lects thousands of depth video sequences of indoor scenes for ob-
share and sell 3D-printable models online (e.g., shapeways.com, ject classification [Nathan Silberman & Fergus 2012]. The Prince-
sketchfab.com, thingiverse.com). This wealth of data also echoes ton Shape Benchmark collects 1,814 polygonal models of specific
 collected contextual information and computed geometric proper-
ties of our dataset in detail against two existing datasets: MPZ14
and ShapeNetCore. We demonstrate that these represent two ex-
treme cases in terms geometric quality while our dataset provides
a mixture of geometric qualities reflecting real-world settings. All
data and analysis of our dataset are freely available to the public. To
facilitate exploration and future reuse, we provide an easy-to-use
online query interface (see Figure 2). This interface augments the
Thingiverse front-end with our geometric analysis of each model.

1000K # Vertices Thingi10K

MPZ14
10K

0.1K ShapeNetCore

0% 50% 100%
Figure 2: Our online query interface selects subsets of Thingi10K.
# Components
1000
objects (e.g., animals, furniture) from various internet sources for
shape classification [Shilane et al. 2004]. More recently, ShapeNet
collects more than three million annotated models [Chang et al. 10
2015]. The ShapeNetCore subset contains 57,459 single-object
models with semi-automatically generated category information.
Although models from these datasets resemble physical objects, 0% 50% 100%
their geometric characteristics suggest their intention was for vi-
Figure 3: Percentile plots of vertex and component count.
sualization rather than fabrication. These datasets are not suitable
for testing 3D printing techniques.
In addition to generic datasets, a variety of specialized datasets ex- 2 Methodology
ist. For example, Lim et al. provide 219 IKEA 3D models for pose-
estimation [Lim et al. 2013]. Recently, Choi et al. released a dataset Instead of our hiring professional modelers or scanning physical ob-
of 10,000 scanned objects, with a subset of 383 successfully recon- jects, we leverage the availability of 3D models hosted and shared
structed 3D models [Choi et al. 2016]. The Shape Retrieval Contest online. Among all 3D shape repositories, we select Thingiverse
releases multiple datasets each year to test retrieval algorithms in- for its large and active user community, its vast collection of print-
cluding generic [Bronstein et al. 2010b; Li et al. 2012], non-rigid validated designs, and its restriction to open-source licenses.
humans [Pickup et al. 2014], sketch-based shapes [Li et al. 2013; Li
et al. 2014], shape correspondences [Bronstein et al. 2010a], facial As one of the largest online shape repositories, Thingiverse hosts
expressions [Nair & Cavallaro 2008; Veltkamp et al. 2011], and more than a million user-uploaded things, 3D designs consisting of
range scans [Dutagaci et al. 2010]. Our Thingi10K dataset com- one or more 3D models (i.e., one or more mesh files). As of October
plements these sources by providing a specialized dataset for 3D 2015, Thingiverse has more than 2 million active users, with 30-40
printing objects. uploads each week and 1.7 million downloads per month [Maker-
Bot 2015]. Thanks to this community, a design is typically not only
We are not the first to utilize Thingiverse models for academic pur- modeled virtually but also fabricated by one or more users, which
poses. To test a rapid prototyping interface, Mueller et al. con- provides invaluable real-world validations.
sider Thingiverse models, but report that meshing artifacts required
manual cleanup before processing [Mueller et al. 2014]. Beyer Our Thingi10K dataset consists of 10,000 models (from 2011
et al. procedurally collect 2,250 models with specific tags from things) systematically culled from Thingiverse via web crawling.
Thingiverse to test a decomposition algorithm [Beyer et al. 2015]. Rather than randomly sample the entire repository, which may con-
Buehler et al. manually sift through 25,000 models from search re- tain bogus models uploaded by inexperienced users or for testing
sults on Thingiverse to identify 363 models as “assistive technolo- purposes, we focus on things featured on Thingiverse. Featured
gies” [Buehler et al. 2015]. Beyond testing a specific routine, these things are entirely and independently selected by Thingiverse staff
works do not analyze low-level geometric characteristics of the col- based on their design, beauty and manufacturability. In a sense,
lected models. These collected datasets are also not publicly avail- these 10,000 models represent a subset of the top-quality designs
able. on Thingiverse. Thingi10K contains every 3D model of every thing
featured by Thingiverse between Sept. 16, 2009 and Nov. 15, 2015.
Contributions. Unlike previous datasets, our Thingi10K dataset
reflects the variety, complexity and (lack of) quality of 3D print-
ing models. It is immediately useful for testing the performance 3 Analysis
of methods for structural analysis [Stava et al. 2012; Zhou et al.
2013; Umetani & Schmidt 2013], shape optimization [Prévost et al. The 10,000-model dataset comes from 2,011 unique things de-
2013; Bächer et al. 2014; Musialski et al. 2015], or solid geometry signed by 1,083 unique users, covering a large variety. Nearly all
operations [Zhou et al. 2016]. Due to its specialized nature and cor- models are stored as .stl files (9,956); the rest are .obj (42), .ply (1),
related contextual information, we suspect the dataset is also useful and .off (1). We analyze both geometric and contextual information
for machine learning and data mining algorithms. We compare the of our dataset to illustrate its representational quality and diversity.
Thingi10K: 2.8M vertices Thingi10K MPZ14
ShapeNetCore: g=1371
165K vertices g=65

g=4476

MPZ14:
76K vertices
g=4886 g=79 g=78

Figure 4: Highest resolution models from each dataset. Figure 6: Models with the highest genus from each dataset.

Thingi10K ShapeNetCore No isolated vertices

#C=4574 #C=2673 Thingi10K
MPZ14
ShapeNetCore
No duplicated faces

Orientable

#C=8064 Edge manifold

#C=2840 No degenerated faces

#C=1620 #C=1100

Figure 5: Connected components of Thingi10K models tend to rep- Vertex manifold

resent salient parts; those in ShapeNetCore are often just discon-
nected patches (models with most components shown). No coplanar intersection

Closed

3.1 Geometry information PWN

We analyze a variety of mesh complexity and quality measures on No self−intersection

our dataset of 3D printing models and compare with two existing
datasets: MPZ14 (116 models) and ShapeNetCore (2000 models Solid
uniformly sampled from 57,459).
0% 25% 50% 75% 100%
3.1.1 Complexity
Figure 7: MPZ14 models are “too clean,” whereas ShapeNetCore
Complexity of 3D model does not directly correlate with 3D print- are unrealistically corrupted in the context of 3D printing.
ing cost. We evaluated three different measures to quantify the com-
plexity of our dataset: number of vertices, number of disconnected
components and genus.
highest genus over 60 times larger than in MPZ14 (see Figure
Figure 3 provides the percentile plot of both vertex and component 6). To avoid confusion, we limit the genus comparison to single-
count over each dataset. The vertex count plot indicates that the component, closed and manifold meshes. Zero of the ShapeNet-
MPZ14 dataset favors moderately high resolution models and ex- Core models meet this criteria.
cludes extremely low or high resolution models. On the other hand,
the distribution over our dataset and ShapeNetCore is similar, with 3.1.2 Mesh quality
our dataset covering a larger range. Figure 4 illustrates the highest
resolution model of each dataset. Mesh qualities of a dataset play a major role in determining its us-
Many geometry processing algorithms assume input will be pro- ability and representation of models in the wild. For example, de-
cessed one component at a time, so it is not a surprise that MPZ14 generate or sliver triangles will cause poor accuracy in non-robust
contains exclusively single-component models. This assumption finite element simulations, and fragile volumetric meshing routines
is not valid in the context of 3D printing, where multiple compo- will fail in the presence of self-intersections. It is crucial to under-
nents could overlap to form a larger shape. Analysing each com- stand the mesh quality of real-world input data in order to design
ponent separately may lead to incorrect results. Within our dataset robust and practical algorithms. Existing datasets often focus on
29% of models have more than one component. ShapeNetCore is high-level properties and provide little insight on their mesh quali-
83% multi-component, but close inspection finds many models are ties. Our analysis aims to fill this gap.
composed of incoherent patches or isolated faces (see Figure 5) In We analyze 13 mesh quality measurements:
contrast, 3D printing models with high numbers of components in
Thingi10K are typically by design, with the base shape naturally Closed: Every edge is adjacent to 2 or more faces.
decomposing into smaller components.
Oriented: Every non-boundary edge has zero signed incidence. In
The genus distribution of our Thingi10K dataset is similar to other words, the number of positively oriented incident faces must
MPZ14, but our dataset covers a larger range of genus, with the equal to the number of negatively oriented incident faces.
No isolated vertices: All vertices are adjacent to at least one face. mesh quality issues not common to 3D printed models, reflecting
that it is gathered from a larger space of 3D models.
No duplicated faces: There does not exist a pair of faces sharing
the same set of vertices. MPZ14 has perfect mesh quality according to seven different
measures. In particular, all models are manifold, oriented and
Vertex-manifold: The one-ring neighborhood of every vertex is a degeneracy-free. Because many geometry processing algorithms
topological disc. do not require the input model to be closed or self-intersection free,
Edge-manifold: Every non-boundary edge must be incident to ex- data from MPZ14 are perfect as proof-of-concept examples. How-
actly two faces. ever, their high quality is due to the fact that models were selected
not on merits of their shape, functionality or aesthetics, but rather
No degeneracy: All faces must have non-collinear vertices. De- because they meet certain quality criteria or have been sanitized.
generacy can be checked with exact predicates [Shewchuk 1997].
On the other hand, ShapeNetCore has very poor mesh quality ac-
No self-intersection: The intersection of any two faces is either cording to 6 measures in Figure 7. Its maximum and average tri-
empty, a shared vertex, or a shared edge. Exact predicates are nec- angle aspect ratios are visibly worse than Thingi10K and MPZ14.
essary to ensure correctness. This is partially due to the fact that these data are collected di-
rectly from the internet, where models were not necessary designed
No coplanar intersections: No two faces are coplanar and overlap- for fabrication purposes. Many existing learning algorithms side-
ping. This is a strictly weaker condition than “no self-intersection.” step the quality issues by transforming boundary representations to
Piecewise-constant winding number (PWN): The winding num- depth images or bounding box hierarchies [Hu 2012]. Performing
ber field at any non-mesh point is piece-wise constant ([Zhou et al. geometry processing algorithms directly on these models is very
2016]). hard due to poor mesh quality.

Solid: The input mesh must be a valid boundary of a subspace of In contrast, our dataset offers a curated collection of 3D meshes
R3 . Specifically, it must be PWN, self-intersection free and induce with a large range of mesh qualities. It contains a significant num-
a {0, 1} winding number field. ber of high quality models as well as a non-negligible proportion
of models with common mesh quality problems. Due to its large
Aspect ratio: The aspect ratio of a triangle is the ratio of its cir- quantity, our dataset is ideal for stress-testing purposes where one
cumradius to the diameter of its incircle. can easily select a subset of the data that matches any combination
of mesh criteria (Section 4). Because all data are sampled from
Intrinsically Delaunay: All edges must have non-negative cotan- real-world models designed to be 3D printed, our dataset provides
gent weights [Fisher et al. 2007]. an unbiased view of the mesh qualities used in practice. Our anal-
These mesh quality measures are not by no means complete. Ad- ysis could be used to gauge the restrictions posed by various as-
ditional quality measures ([Shewchuk 2002; Attene 2013]) can be sumptions on mesh quality. For example, an algorithm assuming
easily adopted. self-intersection-free input would automatically exclude 45% of in-
puts, which may not be acceptable in a real-world settings.

1015 Max aspect ratio

Thingi10K 3.2 Contextual information
1011
ShapeNetCore Each thing in our dataset is annotated by its original designer. Thin-
107
giverse supports three types of annotations: category, subcategory
103 MPZ14 and tags. The first two must be selected from a predefined list of
categories, and the last one is a set of free-form texts created by the
0% 50% 100% user. A total of 4892 distinct tags are used in our dataset. Figure 10
Average aspect ratio illustrates the most frequently used tags.
1011
Unlike ShapeNet [Chang et al. 2015], which focuses on providing
107 categorical annotations specific to object classification purposes,
our dataset comes with a rich and diverse set of original tags rang-
103 ing from the semantics of a 3D model to the printer/material used
for fabrication. For example, Figure 9 shows all models with tags
0% 50% 100%
math, sculpture and scan.
30% Non-intrinsic Delaunay Edges When combined with geometric analysis, our annotations reveal in-
20% teresting insights unavailable from previous works. For example, a
simple frequency analysis indicates OpenSCAD is the most pop-
10% ular modeling tool used by Thingiverse users. Our dataset shows
that 98% of OpenSCAD models are closed, while only 91% of
SketchUp models and 85% of TinkerCAD models are closed
0% 50% 100%
Furthermore, due its fabrication-focused nature, many uploaded
Figure 8: Percentile plots mesh quality measures. meshes are “print-ready” in the sense that their orientation and de-
compositions are designed for optimal printing outcome (See figure
Figure 7 shows the percentage of models that satisfy each of the first 11). Recent papers have tried to solve problems such as decompos-
11 quality measures. Figure 8 illustrates the maximum, average as- ing a large model to fit in the print volume and finding ideal print
pect ratio and the fraction of non-intrinsic Delaunay edges over all orientations [Chen et al. 2015]. The Thingi10K models, by their
models in each dataset. Our analysis shows that MPZ14 has “un- intrinsic nature of being successful prints, represent ground truth
realistically pristine” mesh quality, whereas ShapeNetCore exhibits data.
 Figure 9: Models with tag math (left), sculpture (middle) and scan (right).

modelgametoy
customizer_challenge auto-generated python script to batch download results for custom
sketchup halloween search terms.
plastic_valley supportless education
gears camera reprap fun customizer skull
art
Users of our online query interface can view all contextual and
sculpture rpg robotics
scan vase toolkitchen household geometry model details (Figure 13). In particular, we respect the
space iphone fantasy printbot puzzle playset copyright of each model. On the model detail page, we clearly in-
miniature
ultimaker electronics christmas useful container dicate the original author and open source licence of each model.
castle music animal lamp holder pla monster gear We also provide links to the original Thingiverse pages where the
mount robot dualstrusion lulzbot
math experiment geometry box arduino car
raw data can be obtained.

openscad parametric led casetinkercad

123d_catch capturedornament To demonstrate the power of our online query interface, Figure 14
light jewelry makerbot instrument building
modular
replicator newmuseumchallenge architecture
shows some interesting search results and the query used.

Figure 10: Thingi10K user tags highlight the dataset’s variety. 5 Conclusion

In this work, we present a large-scale annotated 3D dataset based

on models used in 3D printing applications. Our dataset consists of
10,000 meshes crawled systematically from Thingiverse. We ana-
lyze both the contextual and geometric information of our dataset
+
and compare with two existing 3D model datasets. Our analysis
= shows our data covers a large range of categories and provides a
balanced representation of real-world data in terms of mesh com-
plexity and quality. The entire dataset and our analysis are freely
available to the public, and we provide a query interface to facilitate
the exploration and dissection of our dataset.

Figure 11: A soap bubble chair is decomposed and re-oriented by Our dataset could be used as input for stress-testing purposes as
its designer for support-free 3D printing. well as ground truth for learning algorithms. As for future work,
we plan to update and increase the size of the dataset over time
to reflect the fast-evolving nature of the 3D printing community.
Specifically, we would like to include all featured things from Thin-
Lastly, all things are published under one of the open source li- giverse and add support for users to suggest additional models for
censes. Figure 12 illustrates all licenses supported by Thingiverse. inclusion. We hope our dataset and the accompanying analysis pro-
vide an informative summary of 3D printing models and clarify the
4 Online query interface requirements for geometry processing algorithms to be robust.

To facilitate our goal of understanding 3D printed shapes, we pro-

vide an online query interface, ten-thousand-models.appspot.com A Attribution
for anyone to explore and dissect the dataset. The query terms CC Creative Commons CCASa
may consist of one or more clauses. Each clause specifies a single Nc Non-commercial
search condition, e.g. “genus>100”. Multiple clauses are separated Nd No Derivatives LGPL
by commas, and the search engine retrieves models that satisfy all
Pd Public Domain BSD
search conditions.
Sa Share alike CCANd
CCA
Our query interface is very useful in dissecting the dataset based on Pd
mesh quality measures. For example, all single-component, mani- CCPd
fold solid meshes without self-intersection and degeneracies can be GPL CCANc
obtained with the query term “num component=1, is man- ANcNd
ifold, is solid, without self-intersection, ANcSa
without degeneracy”. All meshes satisfying these criteria
are listed on the result page (Figure 2). We also provide an Figure 12: All 10,000 models come under open source licenses.
Acknowledgments D UTAGACI , H., G ODIL , A., C HEUNG , C. P., F URUYA , T., H IL -
LENBRAND , U., AND O HBUCHI , R. 2010. Shrec’10 track:
We thank M. Campen, C. Tymms, and J. Panetta for early feedback Range scan retrieval. In 3DOR.
and proofreading. Funded in part by NSF grants CMMI-11-29917,
IIS-14-09286, and IIS-17257. F ISHER , M., S PRINGBORN , B., S CHRÖDER , P., AND B OBENKO ,
A. I. 2007. An algorithm for the construction of intrinsic delau-
nay triangulations with applications to digital geometry process-
References ing. Computing 81, 2-3, 199–213.
H U , W. 2012. Learning 3d object templates by hierarchical quan-
A IM @S HAPE, 2004. Aim@shape digital shape workbench v5.0 -
tization of geometry and appearance spaces. In CVPR.
shape repository. http://visionair.ge.imati.cnr.it.
L EVOY, M., G ERTH , J., C URLESS , B., AND P ULL ,
ATTENE , M. 2013. Surface mesh qualities.
K., 2005. The stanford 3d scanning repository.
BÄCHER , M., W HITING , E., B ICKEL , B., AND S ORKINE - http://graphics.stanford.edu/data/3Dscanrep/.
H ORNUNG , O. 2014. Spin-it: Optimizing moment of inertia
L I , B., G ODIL , A., AONO , M., BAI , X., F URUYA , T., L I , L.,
for spinnable objects. ACM Trans. Graph..
L ÓPEZ -S ASTRE , R. J., J OHAN , H., O HBUCHI , R., R EDONDO -
B EYER , D., G UREVICH , S., M UELLER , S., C HEN , H.-T., AND C ABRERA , C., ET AL . 2012. Shrec’12 track: Generic 3d shape
BAUDISCH , P. 2015. Platener: Low-fidelity fabrication of 3d retrieval. In 3DOR.
objects by substituting 3d print with laser-cut plates. In Proc. L I , B., L U , Y., G ODIL , A., S CHRECK , T., AONO , M., J OHAN ,
SIGCHI. H., S AAVEDRA , J. M., AND TASHIRO , S. 2013. Shrec’13
B RONSTEIN , A., B RONSTEIN , M., C ASTELLANI , U., D UBROV- track: large scale sketch-based 3d shape retrieval. In EW3DOR.
INA , A., G UIBAS , L., H ORAUD , R., K IMMEL , R., K NOSSOW,
L I , B., L U , Y., L I , C., G ODIL , A., S CHRECK , T., AONO , M.,
D., VON L AVANTE , E., M ATEUS , D., ET AL . 2010. Shrec B URTSCHER , M., F U , H., F URUYA , T., J OHAN , H., ET AL .
2010: robust correspondence benchmark. In EW3DOR. 2014. Shrec’14 track: extended large scale sketch-based 3d
B RONSTEIN , A., B RONSTEIN , M., C ASTELLANI , U., FALCI - shape retrieval. In EW3DOR.
DIENO , B., F USIELLO , A., G ODIL , A., G UIBAS , L., KOKKI -
L IM , J. J., P IRSIAVASH , H., AND T ORRALBA , A. 2013. Parsing
NOS , I., L IAN , Z., OVSJANIKOV, M., ET AL . 2010. Shrec 2010:
IKEA Objects: Fine Pose Estimation. ICCV.
robust large-scale shape retrieval benchmark. Proc. 3DOR.
M AKER B OT, 2015. Celebrating a maker milestone: 1 million up-
B UEHLER , E., B RANHAM , S., A LI , A., C HANG , J. J., H OF - loads on makerbot’s thingiverse .
MANN , M. K., H URST, A., AND K ANE , S. K. 2015. Sharing http://www.makerbot.com/blog/2015/10/29.
is caring: Assistive technology designs on thingiverse. In Proc.
SIGCHI. M UELLER , S., M OHR , T., G UENTHER , K., F ROHNHOFEN , J.,
AND BAUDISCH , P. 2014. faBrickation: Fast 3d printing
C HANG , A. X., F UNKHOUSER , T., G UIBAS , L., H ANRA - of functional objects by integrating construction kit building
HAN , P., H UANG , Q., L I , Z., S AVARESE , S., S AVVA , M., blocks. In Proc. SIGCHI.
S ONG , S., S U , H., X IAO , J., Y I , L., AND Y U , F. 2015.
ShapeNet: An Information-Rich 3D Model Repository. Tech. M USIALSKI , P., AUZINGER , T., B IRSAK , M., W IMMER , M.,
Rep. arXiv:1512.03012 [cs.GR]. AND KOBBELT, L. 2015. Reduced-order shape optimization
using offset surfaces. ACM Trans. Graph..
C HEN , X., Z HANG , H., L IN , J., H U , R., L U , L., H UANG , Q.,
B ENES , B., C OHEN -O R , D., AND C HEN , B. 2015. Dapper: M YLES , A., P IETRONI , N., AND Z ORIN , D. 2014. Robust field-
Decompose-and-pack for 3d printing. ACM Trans. Graph.. aligned global parametrization. ACM Trans. Graph..
C HOI , S., Z HOU , Q.-Y., M ILLER , S., AND KOLTUN , V. 2016. A NAIR , P., AND C AVALLARO , A. 2008. Shrec’08 entry: registration
large dataset of object scans. arXiv:1602.02481. and retrieval of 3d faces using a point distribution model.

Figure 13: Both contextual and geometric information of each

model are available on its model detail page.
 genus>100 num components>100

tag:tinkercad, # component>1 tag:sketchup, is solid tag:math, genus>=1 tag:openscad, is not closed

tag:sculpture, genus>1 tag:scan, is closed

not orientable, num faces>50000

is solid, is not manifold

Figure 14: Our web interface returns subsets of the Thingi10K dataset via text queries.
NATHAN S ILBERMAN , D EREK H OIEM , P. K., AND F ERGUS , R.
2012. Indoor segmentation and support inference from rgbd im-
ages. In ECCV.
P ICKUP, D., S UN , X., ROSIN , P. L., M ARTIN , R., C HENG ,
Z., L IAN , Z., AONO , M., H AMZA , A. B., B RONSTEIN , A.,
B RONSTEIN , M., ET AL . 2014. Shrec’14 track: Shape retrieval
of non-rigid 3d human models. In EW3DOR.
P RÉVOST, R., W HITING , E., L EFEBVRE , S., AND S ORKINE -
H ORNUNG , O. 2013. Make it stand: Balancing shapes for 3d
fabrication. ACM Trans. Graph..
S HEWCHUK , J. R. 1997. Adaptive precision floating-point arith-
metic and fast robust geometric predicates. Discrete & Compu-
tational Geometry 18, 3, 305–363.
S HEWCHUK , J. 2002. What is a good linear finite element?
interpolation, conditioning, anisotropy, and quality measures
(preprint).
S HILANE , P., M IN , P., K AZHDAN , M., AND F UNKHOUSER , T.
2004. The princeton shape benchmark. In Shape modeling ap-
plications, 2004. Proceedings, IEEE, 167–178.
S TAVA , O., VANEK , J., B ENES , B., C ARR , N., AND M ĚCH , R.
2012. Stress relief: Improving structural strength of 3d printable
objects. ACM Trans. Graph..
T URK , G., 2000. The stanford bunny .
http://www.cc.gatech.edu/~turk/bunny/bunny.html.
U METANI , N., AND S CHMIDT, R. 2013. Cross-sectional struc-
tural analysis for 3d printing optimization. In SIGAsia Technical
Briefs.
V ELTKAMP, R. C., VAN J OLE , S., D RIRA , H., A MOR , B. B.,
DAOUDI , M., L I , H., C HEN , L., C LAES , P., S MEETS , D.,
H ERMANS , J., ET AL . 2011. Shrec’11 track: 3d face models
retrieval. In 3DOR.
Z HOU , Q., PANETTA , J., AND Z ORIN , D. 2013. Worst-case struc-
tural analysis. ACM Trans. Graph..
Z HOU , Q., G RINSPUN , E., Z ORIN , D., AND JACOBSON , A. 2016.
Mesh arrangements for solid geometry. ACM Trans. Graph..