ShapeNet: An Information-Rich 3D Model Repository
http://www.shapenet.org
Angel X. Chang1, Thomas Funkhouser2, Leonidas Guibas1, Pat Hanrahan1, Qixing Huang3, Zimo Li3,
Silvio Savarese1, Manolis Savva∗1, Shuran Song2, Hao Su∗1, Jianxiong Xiao2, Li Yi1, and Fisher Yu2
1Stanford University — 2Princeton University — 3Toyota Technological Institute at Chicago
Authors listed alphabetically
Abstract
We present ShapeNet: a richly-annotated, large-scale
repository of shapes represented by 3D CAD models of ob-
jects. ShapeNet contains 3D models from a multitude of
semantic categories and organizes them under the Word-
Net taxonomy. It is a collection of datasets providing many
semantic annotations for each 3D model such as consis-
tent rigid alignments, parts and bilateral symmetry planes,
physical sizes, keywords, as well as other planned anno-
tations. Annotations are made available through a pub-
lic web-based interface to enable data visualization of ob-
ject attributes, promote data-driven geometric analysis, and
provide a large-scale quantitative benchmark for research
in computer graphics and vision. At the time of this techni-
cal report, ShapeNet has indexed more than 3,000,000 mod-
els, 220,000 models out of which are classified into 3,135
categories (WordNet synsets). In this report we describe the
ShapeNet effort as a whole, provide details for all currently
available datasets, and summarize future plans.
1. Introduction
Recent technological developments have led to an ex-
plosion in the amount of 3D data that we can generate and
store. Repositories of 3D CAD models are expanding con-
tinuously, predominantly through aggregation of 3D content
on the web. RGB-D sensors and other technology for scan-
ning and reconstruction are providing increasingly higher
fidelity geometric representations of objects and real envi-
ronments that can eventually become CAD-quality models.
At the same time, there are many open research prob-
lems due to fundamental challenges in using 3D content.
Computing segmentations of 3D shapes, and establishing
correspondences between them are two basic problems in
geometric shape analysis. Recognition of shapes from par-
∗Contact authors: {msavva,haosu}@cs.stanford.edu
tial scans is a research goal shared by computer graphics
and vision. Scene understanding from 2D images is a grand
challenge in vision that has recently benefited tremendously
from 3D CAD models [28, 34]. Navigation of autonomous
robots and planning of grasping manipulations are two large
areas in robotics that benefit from an understanding of 3D
shapes.
At the root of all these research problems lies
the need for attaching semantics to representations of 3D
shapes, and doing so at large scale.
Recently, data-driven methods from the machine learn-
ing community have been exploited by researchers in vision
and NLP (natural language processing). “Big data” in the
visual and textual domains has led to tremendous progress
towards associating semantics with content in both fields.
Mirroring this pattern, recent work in computer graphics
has also applied similar approaches to specific problems in
the synthesis of new shape variations [10] and new arrange-
ments of shapes [6]. However, a critical bottleneck facing
the adoption of data-driven methods for 3D content is the
lack of large-scale, curated datasets of 3D models that are
available to the community.
Motivated by the far-reaching impact of dataset efforts
such as the Penn Treebank [20], WordNet [21] and Ima-
geNet [4], which collectively have tens of thousands of ci-
tations, we propose establishing ShapeNet: a large-scale 3D
model dataset. Making a comprehensive, semantically en-
riched shape dataset available to the community can have
immense impact, enabling many avenues of future research.
In constructing ShapeNet we aim to fulfill several goals:
• Collect and centralize 3D model datasets, helping to
organize effort in the research community.
• Support data-driven methods requiring 3D model data.
• Enable evaluation and comparison of algorithms for
fundamental tasks involving geometry (e.g., segmen-
tation, alignment, correspondence).
• Serve as a knowledge base for representing real-world
objects and their semantics.
1
arXiv:1512.03012v1  [cs.GR]  9 Dec 2015
These goals imply several desiderata for ShapeNet:
• Broad and deep coverage of objects observed in the
real world, with thousands of object categories and
millions of total instances.
• Categorization scheme connected to other modalities
of knowledge such as 2D images and language.
• Annotation of salient physical attributes on models,
such as canonical orientations, planes of symmetry,
and part decompositions.
• Web-based interfaces for searching, viewing and re-
trieving models in the dataset through several modali-
ties: textual keywords, taxonomy traversal, image and
shape similarity search.
Achieving these goals and providing the resulting dataset
to the community will enable many advances and applica-
tions in computer graphics and vision.
In this report, we first situate ShapeNet, explaining the
overall goals of the effort and the types of data it is in-
tended to contain, as well as motivating the long-term vi-
sion and infrastructural design decisions (Section 3). We
then describe the acquisition and validation of annotations
collected so far (Section 4), summarize the current state of
all available ShapeNet datasets, and provide basic statistics
on the collected annotations (Section 5). We end with a dis-
cussion of ShapeNet’s future trajectory and connect it with
several research directions (Section 7).
2. Background and Related Work
There has been substantial growth in the number of of
3D models available online over the last decade, with repos-
itories like the Trimble 3D Warehouse providing millions
of 3D polygonal models covering thousands of object and
scene categories. Yet, there are few collections of 3D mod-
els that provide useful organization and annotations. Mean-
ingful textual descriptions are rarely provided for individ-
ual models, and online repositories are usually either un-
organized or grouped into gross categories (e.g., furniture,
architecture, etc. [7]). As a result, they have been poorly
utilized in research and applications.
There have been previous efforts to build organized col-
lections of 3D models (e.g., [5, 7]). However, they have
provided quite small datasets, covered only a small num-
ber of semantic categories, and included few structural and
semantic annotations. Most of these previous collections
have been developed for evaluating shape retrieval and clas-
sification algorithms. For example, datasets are created an-
nually for the Shape Retrieval Contest (SHREC) that com-
monly contains sets of models organized in object cate-
gories. However, those datasets are very small — the most
recent SHREC iteration in 2014 [17] contains a “large”
dataset with around 9,000 models consisting of models from
a variety of sources organized into 171 categories (Table 1).
The Princeton Shape Benchmark is probably the most
well-known and frequently used 3D shape collection to date
(with over 1000 citations) [27]. It contains around 1,800 3D
models grouped into 90 categories, but has no annotations
beyond category labels.
Other commonly-used datasets
contain segmentations [2], correspondences [13, 12], hier-
archies [19], symmetries [11], salient features [3], seman-
tic segmentations and labels [36], alignments of 3D models
with images [35], semantic ontologies [5], and other func-
tional annotations — but again only for small size datasets.
For example, the Benchmark for 3D Mesh Segmentation
contains just 380 models in 19 object classes [2].
In contrast, there has been a flurry of activity on collect-
ing, organizing, and labeling large datasets in computer vi-
sion and related fields. For example, ImageNet [4] provides
a set of 14M images organized into 20K categories asso-
ciated with “synsets” of WordNet [21]. LabelMe provides
segmentations and label annotations of hundreds of thou-
sands of objects in tens of thousands of images [24]. The
SUN dataset provides 3M annotations of objects in 4K cat-
egories appearing in 131K images of 900 types of scenes.
Recent work demonstrated the benefit of a large dataset
of 120K 3D CAD models in training a convolutional neu-
ral network for object recognition and next-best view pre-
diction in RGB-D data [34]. Large datasets such as this
and others (e.g., [14, 18]) have revitalized data-driven al-
gorithms for recognition, detection, and editing of images,
which have revolutionized computer vision.
Similarly, large collections of annotated 3D data have
had great influence on progress in other disciplines. For ex-
ample, the Protein Data Bank [1] provides a database with
100K protein 3D structures, each labeled with its source
and links to structural and functional annotations [15]. This
database is a common repository of all 3D protein structures
solved to date and provides a shared infrastructure for the
collection and transfer of knowledge about each entry. It has
accelerated the development of data-driven algorithms, fa-
cilitated the creation of benchmarks, and linked researchers
and industry from around the world. We aim to provide a
similar resource for 3D models of everyday objects.
3. ShapeNet: An Information-Rich 3D Model
Repository
ShapeNet is a large, information-rich repository of 3D
models. It contains models spanning a multitude of seman-
tic categories. Unlike previous 3D model repositories, it
provides extensive sets of annotations for every model and
links between models in the repository and other multime-
dia data outside the repository.
Like ImageNet, ShapeNet provides a view of the con-
tained data in a hierarchical categorization according to
WordNet synsets (Figure 1). Unlike other model reposi-
tories, ShapeNet also provides a rich set of annotations for
2
Benchmarks
Types
# models
# classes
Avg # models per class
SHREC14LSGTB
Generic
8,987
171
53
PSB
Generic
907+907 (train+test)
90+92 (train+test)
10+10 (train+test)
SHREC12GTB
Generic
1200
60
20
TSB
Generic
10,000
352
28
CCCC
Generic
473
55
9
WMB
Watertight (articulated)
400
20
20
MSB
Articulated
457
19
24
BAB
Architecture
2257
183+180 (function+form)
12+13 (function+form)
ESB
CAD
867
45
19
Table 1. Source datasets from SHREC 2014:
Princeton Shape Benchmark (PSB) [27], SHREC 2012 generic Shape Benchmark
(SHREC12GTB) [16], Toyohashi Shape Benchmark (TSB) [29], Konstanz 3D Model Benchmark (CCCC) [32], Watertight Model Bench-
mark (WMB) [31], McGill 3D Shape Benchmark (MSB) [37], Bonn Architecture Benchmark (BAB) [33], Purdue Engineering Shape
Benchmark (ESB) [9].
each shape and correspondences between shapes. The an-
notations include geometric attributes such as upright and
front orientation vectors, parts and keypoints, shape sym-
metries (reflection plane, other rotational symmetries), and
scale of object in real world units. These attributes provide
valuable resources for processing, understanding and visu-
alizing 3D shapes in a way that is aware of the semantics of
the shape.
We have currently collected approximately 3 million
shapes from online 3D model repositories, and categorized
300 thousand of them against the WordNet taxonomy. We
have also annotated a subset of these models with shape
properties such as upright and front orientations, symme-
tries, and hierarchical part decompositions. We are contin-
uing the process of expanding the annotated set of models
and also collecting new models from new data sources.
In the following sections, we discuss how 3D models
are collected for ShapeNet, what annotations will be added,
how those annotations will be generated, how annotations
will be updated as the dataset evolves over time, and what
tools will be provided for the community to search, browse,
and utilize existing data, as well as contribute new data.
3.1. Data Collection
The raw 3D model data for ShapeNet comes from public
online repositories or existing research datasets. ShapeNet
is intended to be an evolving repository with regular updates
as more and more 3D models become available, as more
people contribute annotations, and as the data captured with
new 3D sensors become prevalent.
We have collected 3D polygonal models from two
popular public repositories: Trimble 3D Warehouse1 and
Yobi3D2. The Trimble 3D Warehouse contains 2.4M user-
designed 3D models and scenes. Yobi3D contains 350K
additional models collected from a wide range of other on-
line repositories. Together, they provide a diverse set of
1https://3dwarehouse.sketchup.com/
2https://yobi3d.com
Figure 1. Screenshot of the online ShapeNet taxonomy view, or-
ganizing contained 3D models under WordNet synsets.
shapes from a broad set of object and scene categories
— e.g., many organic shape categories (e.g., humans and
mammals), which are rare in Warehouse3D, are plentiful in
Yobi3D. For more detailed statistics on the currently avail-
able ShapeNet models refer to Section 5.
Though the tools developed for this project will be
general-purpose, we intend to include only 3D models of
objects encountered by people in the everyday world. That
is, it will not include CAD mechanical parts, molecular
structures, or other domain-specific objects. However, we
will include scenes (e.g., office), objects (e.g., laptop com-
puter), and parts of objects (e.g., keyboard). Models are
organized under WordNet [21] noun “synsets” (synonym
sets). WordNet provides a broad and deep taxonomy with
over 80K distinct synsets representing distinct noun con-
cepts arranged as a DAG network of hyponym relationships
(e.g., “canary” is a hyponym of “bird”). This taxonomy has
been used by ImageNet to describe categories of objects at
3
multiple scales [4]. Though we first use WordNet due to its
popularity, the ShapeNet UI is designed to allow multiple
views into the collection of shapes that it contains, includ-
ing different taxonomy views and faceted navigation.
3.2. Annotation Types
We envision ShapeNet as far more than a collection of
3D models. ShapeNet will include a rich set of annota-
tions that provide semantic information about those mod-
els, establish links between them, and links to other modal-
ities of data (e.g., images). These annotations are exactly
what make ShapeNet uniquely valuable. Figure 2 illustrates
the value of this dense network of interlinked attributes on
shapes, which we describe below.
Language-related
Annotations:
Naming
objects
by
their basic category is useful for indexing, grouping, and
linking to related sources of data. As described in the pre-
vious section, we organize ShapeNet based on the Word-
Net [21] taxonomy. Synsets are interlinked with various
relations, such as hyper and hyponym, and part-whole rela-
tions. Due to the popularity of WordNet, we can leverage
other resources linked to WordNet such as ImageNet, Con-
ceptNet, Freebase, and Wikipedia. In particular, linking to
ImageNet [4] will help transport information between im-
ages and shapes. We assign each 3D model in ShapeNet to
one or more synsets in the WordNet taxonomy (i.e., we pop-
ulate each synset with a collection of shapes). Please refer
to Section 4.1 for details on the acquisition and validation
of basic category annotations. Future planned annotations
include natural language descriptions of objects and object
part-part relation descriptions.
Geometric Annotations:
A critical property that distin-
guishes ShapeNet from image and video datasets is the fi-
delity with which 3D geometry represents real-world struc-
tures.
We combine algorithmic predictions and manual
annotations to organize shapes by category-level geomet-
ric properties and further derive rich geometric annotations
from the raw 3D model geometry.
• Rigid Alignments: Establishing a consistent canon-
ical orientation (e.g., upright and front) for every
model is important for various tasks such as visual-
izing shapes [13], shape classification [8] and shape
recognition [34]. Fortunately, most raw 3D model data
is by default placed in an upright orientation, and the
front orientations are typically aligned with an axis.
This allows us to use a hierarchical clustering and
alignment approach to ensure consistent rigid align-
ments within each category (see Section 4.2).
• Parts and Keypoints: Many shapes contain or have
natural decompositions into important parts, as well as
significant keypoints related to both their geometry and
their semantics. For example, often different materials
are associated with different parts. We intend to cap-
ture as much of that as possible into ShapeNet.
• Symmetry: Bilateral symmetry planes and rotational
symmetries are prevalent in artificial and natural ob-
jects, and deeply connected with the alignment and
functionality of shapes. We refer to Section 4.4 for
more details on how we compute symmetries for the
shapes in ShapeNet.
• Object Size: Object size is useful for many applica-
tions, such as reducing the hypothesis space in object
recognition.
Size annotations are discussed in Sec-
tion 5.2.
Functional Annotations:
Many objects, especially man-
made artifacts such as furniture and appliances, can be used
by humans. Functional annotations describe these usage
patterns. Such annotations are often highly correlated with
specific regions of an object. In addition, it is often related
with the specific type of human action. ShapeNet aims to
store functional annotations at the global shape level and at
the object part level.
• Functional Parts: Parts are critical for understand-
ing object structure, human activities involving a 3D
shape, and ergonomic product design. We plan to an-
notate parts according to their function — in fact the
very definition of parts has to be based on both geo-
metric and functional criteria.
• Affordances: We are interested in affordance annota-
tions that are function and activity specific. Examples
of such annotations include supporting plane annota-
tions, and graspable region annotations for various ob-
ject manipulations.
Physical Annotations:
Real objects exist in the physical
world and typically have fixed physical properties such as
dimensions and densities.
Thus, it is important to store
physical attribute annotations for 3D shapes.
• Surface Material: We are especially interested in the
optical properties and semantic names of surface mate-
rials. They are important for applications such as ren-
dering and structural strength estimation.
• Weight: A basic property of objects which is very use-
ful for physical simulations, and reasoning about sta-
bility and static support.
In general, the issue of compact and informative rep-
resentations for all the above attributes over shapes raises
many interesting questions that we will need to address
as part of the ShapeNet effort. Many annotations are cur-
rently ongoing projects and involve interesting open re-
search problems.
4
Swivel chair
Backrest
Seat
Base
Leg
Wheel
ImageNet
WordNet synset
Part Hierarchy
Part Correspondences
Link to WordNet Taxonomy
Swivel chair: a chair that swivels 
on its base
 Hypernyms: chair > seat > furniture > ...
 Part meronyms: backrest, seat, base
 Sister terms: armchair, barber chair, ...
Alignment+Symmetry
Dim: 50 x 45 x 5 cm
Material: foam, fabric
Mass: 5 Kg
Function: support
Figure 2. ShapeNet annotations illustrated for an example chair model. Left: links to the WordNet taxonomy provide definitions of objects,
is-a and has-a relations, and a connection to images from ImageNet. Middle-left: shape is aligned to a consistent upright and front
orientation, and symmetries are computed Middle-right: hierarchical decomposition of shape into parts on which various attributes are
defined: names, symmetries, dimensions, materials, and masses. Right: part-to-part and point-to-point connections are established at all
levels within ShapeNet producing a dense and semantically rich network of correspondences. The gray background indicates annotations
that are currently ongoing and not yet available for release.
3.3. Annotation Methodology
Though at first glance it might seem reasonable to collect
the annotations we describe purely through manual human
effort, we will in general take a hybrid approach. For anno-
tation types where it is possible, we will first algorithmically
predict the annotation for each model instance (e.g., global
symmetry planes, consistent rigid alignments). We will then
verify these predictions through crowd-sourcing pipelines
and inspection by human experts. This hybrid strategy is
sensible in the context of 3D shape data as there are already
various algorithms we can leverage, and collecting corre-
sponding annotations entirely through manual effort can be
extremely labor intensive. In particular, since objects in a
3D representation are both more pure and more complete
than objects in images, we can expect better and easier to
establish correspondences between 3D shapes, enabling al-
gorithmic transport of semantic annotations. In many cases,
the design of the human annotation interfaces themselves is
an open question — which stands in contrast to largely man-
ual image labeling efforts such as ImageNet. As a concrete
example, shape part annotation can be presented and per-
formed in various ways with different trade-offs in the type
of obtained part annotation, the accuracy and the efficiency
of the annotation process.
Coupled with this hybrid annotation strategy, we also
take particular care to preserve the provenance and confi-
dence of each algorithmic and human annotation. The anno-
tation source (whether an algorithm, or human effort), and
a measure of the trust we can place in each annotation are
critical pieces of information especially when we have to
combine, aggregate, and reconcile several annotations.
3.4. Annotation Schema and Web API
To provide convenient access to all of the model and an-
notation data contained within ShapeNet, we construct an
index over all the 3D models and their associated annota-
tions using the Apache Solr framework.3 Each stored an-
notation for a given 3D model is contained within the index
as a separate attribute that can be easily queried and filtered
through a simple web-based UI. In addition, to make the
dataset conveniently accessible to researchers, we provide a
batched download capability.
4. Annotation Acquisition and Validation
A key challenge in constructing ShapeNet is the method-
ology for acquiring and validating annotations. Our goal
is to provide all annotations with high accuracy. In cases
where full verification is not yet available, we aim to es-
timate a confidence metric for each annotation, as well as
record its provenance. This will enable others to properly
estimate the trustworthiness of the information we provide
and use it for different applications.
4.1. Category Annotation
As described in Section 3.2, we assign each 3D model to
one or more synsets in the WordNet taxonomy.
Annotation
Models are retrieved by textual query into the
online repositories that we collected, and the initial category
annotation is set to the used textual query for each retrieved
3http://lucene.apache.org/solr/
5
model. After we retrieve these models we use the popular-
ity score of each model on the repository to sort models and
ask human workers to verify the assigned category annota-
tion. This is sensible since the more popular models tend to
be high quality and correctly retrieved through the category
keyword textual query. We stop verifying category annota-
tions with people once the positive ratio is lower than a 2%
threshold.
Clean-up
In order for the dataset to be easily usable by re-
searchers it should contain clean and high quality 3D mod-
els. Through inspection, we identify and group 3D models
into the following categories: single 3D models, 3D scenes,
billboards, and big ground plane.
• Single 3D models: semantically distinct objects; focus
of our ShapeNetCore annotation effort.
• 3D scenes: detected by counting the number of con-
nected components in a voxelized representation. We
manually verify these detections and mark scenes for
future analysis.
• Billboards: planes with a painted texture. Often used
to represent people and trees. These models are gen-
erally not useful for geometric analysis. They can be
detected by checking whether a single plane can fit all
vertices.
• Big ground plane: object of interest placed on a large
horizontal plane or in front of large vertical plane. Al-
though we do not currently use these models, the plane
can easily be identified and removed through simple
geometric analysis.
We currently include the single 3D models in the
ShapeNetCore subset of ShapeNet.
4.2. Hierarchical Rigid Alignment
The goal of this step is to establish a consistent canon-
ical orientation for models within each category.
Such
alignment is important for various tasks such as visualizing
shapes, shape classification and shape recognition. Figure 3
shows several categories in ShapeNet that have been con-
sistently aligned.
Though the concept of consistent orientation seems nat-
ural, one issue has to be addressed. We explain by an ex-
ample. “armchair”, “chair” and “seat” are three categories
in our taxonomy, each being a subcategory of its succes-
sor. Consistent orientation can be well defined for shapes in
the “armchair” category, by checking arms, legs and backs.
Yet, it becomes difficult to define for the “chair” category.
For example, “side chair” and “swivel chair” are both sub-
categories of “chair”, however, swivel chairs have a very
different leg structure than most side chairs. It becomes
even more ambiguous to define for “seat”, which has sub-
categories such as “stool”, “couch”, and “chair”. However,
Figure 3. Examples of aligned models in the chair, laptop, bench,
and airplane synsets.
the concept of an upright orientation still applies throughout
most levels of the taxonomy.
Following the above discussion, it is natural for us to pro-
pose a hierarchical alignment method, with a small amount
of human supervision. The basic idea is to hierarchically
align models following the taxonomy of ShapeNet in a
bottom-up manner, i.e., we start from aligning shapes in
low-level categories and then gradually elevate to higher
level categories. When we proceed to the higher level, the
self-consistent orientation within a subcategory should be
maintained. For the alignment at each level, we first use
a geometric algorithm described in the Appendix A.1, and
then ask human experts to check and correct possible mis-
alignments. With this strategy, we efficiently obtain consis-
tent orientations. In practice, most shapes in the same low-
level categories can be well aligned algorithmically, requir-
ing limited manual correction. Though the proportion of
manual corrections increases for aligning higher-level cate-
gories, the number of categories at each level becomes log-
arithmically smaller.
4.3. Parts and Keypoints
To obtain part and keypoint annotations we start from
some curated part annotations within each category. For
parts, this acquisition can be speeded up by having algo-
rithmically generated segmentations and then having users
accept or modify parts from these. We intend to experiment
with both 2D and 3D interfaces for this task. We then ex-
ploit a number of different algorithmic techniques to propa-
gate this information to other nearby shapes. Such methods
can rely on rigid alignments in 3D, feature descriptor align-
ments in an appropriately defined feature space, or general
shape correspondences. We iterate this pipeline, using ac-
tive learning to estimate the 3D models and regions of these
6
models where further human annotation would be most in-
formative, generate a new set of crowd-sourced annotation
tasks, algorithmically propagate their results, and so on. In
the end we have users verify all proposed parts and key-
points, as verification is much faster than direct annotation.
4.4. Symmetry Estimation
We provide bilateral symmetry plane detections for all
3D models in ShapeNetCore. Our method is a modified
version of [22]. The basic idea is to use hough transform
to vote on the parameters of the symmetry plane. More
specifically, we generate all combinations of pairs of ver-
tices from the mesh. Each pair casts a vote of a possible
symmetry plane in the discretized space of plane parame-
ters partitioned evenly. We then pick the parameter with
the most votes as the symmetry plane candidate. As a final
step, this candidate is verified to ensure that every vertex
has a symmetric counterpart.
4.5. Physical Property Estimation
Before computing physical attribute annotations, the di-
mensions of the models need to be correspond to the real
world. We estimate the absolute dimensions of models us-
ing prior work in size estimation [25], followed by man-
ual verification. With the given absolute dimensions, we
now compute the total solid volume of each model through
filled-in voxelization. We use the space carving approach
implemented by Binvox [23]. Categories of objects that are
known to be container-like (i.e., bottles, microwaves) are
annotated as such and only the surface voxelization volume
is used instead. We then estimate the proportional mate-
rial composition of each object category and use a table of
material densities along with each model instance volume
to compute a rough total weight estimate for that instance.
More details about the acquisition of these physical attribute
annotations are available separately [26].
5. Current Statistics
At the time of this technical report, ShapeNet has in-
dexed roughly 3,000,000 models.
220,000 models of
these models are classified into 3,135 categories (Word-
Net synsets). Below we provide detailed statistics for the
currently annotated models in ShapeNet as a whole, as
well as details of the available publicly released subsets of
ShapeNet.
Category Distribution
Figure 4 shows the distributions
of the number of shapes per synset at various taxonomy
levels for the current ShapeNetCore corpus. To the best of
our knowledge, ShapeNet is the largest clean shape dataset
available in terms of total number of shapes, average num-
ber of shapes per category, as well as the number of cate-
gories.
We observe that ShapeNet as a whole is strongly biased
towards categories of rigid man-made artifacts, due to the
bias of the source 3D model repositories. This is in con-
trast to common image database statistics that contain more
natural objects such as plants and animals [30]. This distri-
bution bias is probably due to a combination of factors: 1)
meshes of natural objects are more difficult to design using
common CAD software; 2) 3D model consumers are typi-
cally more interested in artificial objects such as those ob-
served in modern urban lifestyles. The former factor can be
mitigated in the near future by using the rapidly improving
depth sensing and 3D scanning technology.
5.1. ShapeNetCore
ShapeNetCore is a subset of the full ShapeNet dataset
with single clean 3D models and manually verified category
and alignment annotations. It covers 55 common object cat-
egories with about 51,300 unique 3D models. The 12 object
categories of PASCAL 3D+[35], a popular computer vision
3D benchmark dataset, are all covered by ShapeNetCore.
The category distribution of ShapeNetCore is shown in Ta-
ble 2.
5.2. ShapeNetSem
ShapeNetSem is a smaller, more densely annotated sub-
set consisting of 12,000 models spread over a broader set of
270 categories. In addition to manually verified category
labels and consistent alignments, these models are anno-
tated with real-world dimensions, estimates of their mate-
rial composition at the category level, and estimates of their
total volume and weight. The total numbers of models for
the top 100 categories in this subset are given in Table 3.
6. Discussion and Future Work
The construction of ShapeNet is a continuous, ongoing
effort. Here we have just described the initial steps we have
taken in defining ShapeNet and populating a core subset
of model annotations that we hope will prove useful to the
community. We plan to grow ShapeNet in four distinct di-
rections:
Additional annotation types
We will introduce several
additional types of annotations that have strong connections
to the semantics and functionality of objects. Firstly, hierar-
chical part decompositions of objects will provide a useful
finer granularity description of object structure that can be
leveraged for part segmentation and shape synthesis. Sec-
ondly, physical object property annotations such as materi-
als and their attributes will allow higher fidelity physics and
appearance simulation, adding another layer of understand-
ing to methods in vision and graphics.
7
0K 5K
10K 15K 20K 25K 30K 35K 40K 45K 50K 55K 60K 65K 70K 75K 80K 85K 90K 95K 100K105K110K115K 120K
artifact
plant
person
animal
athletics
natural object
geological
Root
0K
2K
4K
6K
8K
10K
12K
14K
16K
18K
20K
22K
24K
26K
28K
30K
32K
34K
device
furniture
container
transport
equipment
implement
weaponry
Artifact
0K
1K
2K
3K
4K
5K
6K
7K
8K
seat
table
cabinet
lamp
bedroom furniture
bookcase
wardrobe
office furniture
Furniture
0K
1K
2K
3K
4K
5K
motor vehicle
armored vehicle
recreational vehicle
tracked vehicle
trailer
locomotive
railcar
tractor
skateboard
Wheeled Vehicle
0K
1K
2K
3K
4K
5K
6K
chair
sofa
stool
bench
ottoman
toilet seat
Seat
0
500
1000 1500 2000 2500
side chair
armchair
club chair
swivel chair
folding chair
cantilever chair
chaise
rocking chair
lawn chair
rex chair
zigzag chair
wheelchair
barcelona chair
tulip chair
ball chair
Chair
0
100
200
300
400
sedan
sport utility
coupe
racer
sports car
convertible
limousine
minivan
ambulance
roadster
station wagon
touring car
cruiser
jeep
hot rod
hatchback
cab
Car
0
500
1000
1500
2000
2500
desk
rectangular table
workshop table
side table
counter
short table
worktable
Table
Figure 4. Plots of the distribution of ShapeNet models over WordNet synsets at multiple levels of the taxonomy (only the top few children
synsets are shown at each level). The highest level (root) is at the top and the taxonomy levels become lower downwards and to the right.
Note the bias towards rigid man-made artifacts at the top and the broad coverage of many low level categories towards the bottom.
ID
Name
Num
ID
Name
Num
ID
Name
Num
04379243
table
8443
03593526
jar
597
04225987
skateboard
152
02958343
car
7497
02876657
bottle
498
04460130
tower
133
03001627
chair
6778
02871439
bookshelf
466
02942699
camera
113
02691156
airplane
4045
03642806
laptop
460
02801938
basket
113
04256520
sofa
3173
03624134
knife
424
02946921
can
108
04090263
rifle
2373
04468005
train
389
03938244
pillow
96
03636649
lamp
2318
02747177
trash bin
343
03710193
mailbox
94
04530566
watercraft
1939
03790512
motorbike
337
03207941
dishwasher
93
02828884
bench
1816
03948459
pistol
307
04099429
rocket
85
03691459
loudspeaker
1618
03337140
file cabinet
298
02773838
bag
83
02933112
cabinet
1572
02818832
bed
254
02843684
birdhouse
73
03211117
display
1095
03928116
piano
239
03261776
earphone
73
04401088
telephone
1052
04330267
stove
218
03759954
microphone
67
02924116
bus
939
03797390
mug
214
04074963
remote
67
02808440
bathtub
857
02880940
bowl
186
03085013
keyboard
65
03467517
guitar
797
04554684
washer
169
02834778
bicycle
59
03325088
faucet
744
04004475
printer
166
02954340
cap
56
03046257
clock
655
03513137
helmet
162
03991062
flowerpot
602
03761084
microwaves
152
Total
57386
Table 2. Statistics of ShapeNetCore synsets. ID corresponds to WordNet synset offset, which is aligned with ImageNet.
8
Category
Num
Category
Num
Category
Num
Category
Num
Category
Num
Chair
696
Monitor
127
WallLamp
78
Gun
54
FlagPole
38
Lamp
663
RoundTable
120
SideChair
77
Nightstand
53
TvStand
38
ChestOfDrawers
511
TrashBin
117
VideoGameConsole
75
Mug
51
Fireplace
37
Table
427
DrinkingUtensil
112
MediaStorage
73
AccentChair
50
Rack
37
Couch
413
DeskLamp
110
Painting
73
ChessBoard
49
LightSwitch
36
Computer
244
Clock
101
Desktop
71
Rug
49
Oven
36
Dresser
234
ToyFigure
101
AccentTable
70
WallUnit
46
Airplane
35
TV
233
Plant
98
Camera
70
Mirror
45
DresserWithMirror
35
WallArt
222
Armoire
95
Picture
69
Bowl
44
Calculator
34
Bed
221
QueenBed
94
Refrigerator
68
SodaCan
44
TableClock
34
Cabinet
221
Stool
92
Speaker
68
VideoGameController
44
Toilet
34
FloorLamp
201
EndTable
91
Sideboard
67
WallClock
43
Cup
33
Desk
189
Bottle
88
Barstool
66
Printer
42
Stapler
33
PottedPlant
188
DiningTable
88
Guitar
65
Sword
40
PaperBox
32
FoodItem
180
Bookcase
87
MediaPlayer
62
USBStick
40
SpaceShip
32
Laptop
173
CeilingLamp
86
Ipod
59
Chaise
39
Toy
32
Vase
163
Bench
84
PersonStanding
57
OfficeSideChair
39
ToiletPaper
31
TableLamp
142
Book
84
Piano
56
Poster
39
Knife
30
OfficeChair
137
CoffeeTable
81
Curtain
55
Sink
39
PictureFrame
30
CellPhone
130
Pencil
80
Candle
54
Telephone
39
Recliner
30
Table 3. Total number of models for the top 100 ShapeNetSem categories (out of 270 categories). Each category is also linked to the
corresponding WordNet synset, establishing the same linkage to WordNet and ImageNet as with ShapeNetCore.
Correspondences
One of the most important goals of
ShapeNet is to provide a dense network of correspondences
between 3D models and their parts. This will be invalu-
able for enabling much shape analysis research and helping
to improve and evaluate methods for many traditional tasks
such as alignment and segmentation. Additionally, we plan
to provide correspondences between 3D model parts and
image patches in ImageNet — a link that will be critical
for propagating information between image space and 3D
models.
RGB-D data
The rapid proliferation of commodity
RGB-D sensors is already making the process of capturing
real-world environments better and more efficient. Expand-
ing ShapeNet to include shapes reconstructed from scanned
RGB-D data is a critical goal. We foresee that over time,
the amount of available reconstructed shape data will over-
shadow the existing designed 3D model data and as such
this is a natural growth direction for ShapeNet. A related ef-
fort that we are currently undertaking is to align 3D models
to objects observed in RGB-D frames. This will establish
a powerful connection between real world observations and
3D models.
Annotation coverage
We will continue to expand the set
of annotated models to cover a bigger subset of the entirety
of ShapeNet. We will explore combinations of algorithmic
propagation methods and crowd-sourcing for verification of
the algorithmic results.
7. Conclusion
We firmly believe that ShapeNet will prove to be an im-
mensely useful resource to several research communities in
several ways:
Data-driven research
By establishing ShapeNet as the
first large-scale 3D shape dataset of its kind we can help
to move computer graphics research toward a data-driven
direction following recent developments in vision and NLP.
Additionally, we can help to enable larger-scale quantitative
analysis of proposed systems that can clarify the benefits of
particular methodologies against a broader and more repre-
sentative variety of 3D model data.
Training resource
By providing a large-scale, richly an-
notated dataset we can also promote a broad class of re-
cently resurgent machine learning and neural network meth-
ods for applications dealing with geometric data.
Much
like research in computer vision and natural language un-
derstanding, computational geometry and graphics stand
to benefit immensely from these data-driven learning ap-
proaches.
Benchmark dataset
We hope that ShapeNet will grow to
become a canonical benchmark dataset for several evalua-
tion tasks and challenges. In this way, we would like to en-
gage the broader research community in helping us define
and grow ShapeNet to be a pivotal dataset with long-lasting
impact.
References
[1] Helen M Berman, John Westbrook, Zukang Feng, Gary
Gilliland, TN Bhat, Helge Weissig, Ilya N Shindyalov, and
Philip E Bourne. The protein data bank. Nucleic Acids Res,
28:235–242, 2000. 2
[2] Xiaobai
Chen,
Aleksey
Golovinskiy,
and
Thomas
Funkhouser.
A benchmark for 3D mesh segmentation.
ACM TOG, 28(3):73:1–73:12, July 2009. 2
9
[3] Xiaobai Chen, Abulhair Saparov, Bill Pang, and Thomas
Funkhouser. Schelling points on 3D surface meshes. ACM
TOG, August 2012. 2
[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li,
and Li Fei-Fei. ImageNet: A large-scale hierarchical image
database. In CVPR, 2009. 1, 2, 4
[5] Bianca
Falcidieno.
Aim@shape.
http://www.
aimatshape.net/ontologies/shapes/, 2005. 2
[6] Matthew Fisher, Daniel Ritchie, Manolis Savva, Thomas
Funkhouser, and Pat Hanrahan. Example-based synthesis of
3D object arrangements. ACM TOG, 31(6):135, 2012. 1
[7] Paul-Louis George.
Gamma.
http://www.rocq.
inria.fr/gamma/download/download.php,
2007. 2
[8] Qixing Huang, Hao Su, and Leonidas Guibas. Fine-grained
semi-supervised labeling of large shape collections. ACM
TOG, 32:190:1–190:10, 2013. 4, 11
[9] Subramaniam Jayanti, Yagnanarayanan Kalyanaraman, Na-
traj Iyer, and Karthik Ramani. Developing an engineering
shape benchmark for CAD models. Computer-Aided Design,
2006. 3
[10] Evangelos Kalogerakis,
Siddhartha Chaudhuri,
Daphne
Koller, and Vladlen Koltun.
A probabilistic model for
component-based shape synthesis. ACM TOG, 31:55, 2012.
1
[11] Vladimir Kim, Yaron Lipman, Xiaobai Chen, and Thomas
Funkhouser.
Mobius transformations for global intrinsic
symmetry analysis.
Symposium on Geometry Processing,
July 2010. 2
[12] Vladimir G. Kim, Wilmot Li, Niloy J. Mitra, Siddhartha
Chaudhuri, Stephen DiVerdi, and Thomas Funkhouser.
Learning part-based templates from large collections of 3D
shapes. ACM TOG, 32(4):70:1–70:12, July 2013. 2
[13] Vladimir G. Kim, Wilmot Li, Niloy J. Mitra, Stephen
DiVerdi, and Thomas Funkhouser.
Exploring collections
of 3D models using fuzzy correspondences.
ACM TOG,
31(4):54:1–54:11, July 2012. 2, 4
[14] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei.
3D object representations for fine-grained categorization. In
4th International IEEE Workshop on 3D Representation and
Recognition (3dRR-13), Sydney, Australia, 2013. 2
[15] Roman A Laskowski, E Gail Hutchinson, Alex D Michie,
Andrew C Wallace, Martin L Jones, and Janet M Thornton.
PDBsum: A web-based database of summaries and analyses
of all PDB structures. Trends Biochem. Sci., 22:488–490,
1997. 2
[16] Bo Li, Afzal Godil, Masaki Aono, X Bai, Takahiko Furuya,
L Li, R L´opez-Sastre, Henry Johan, Ryutarou Ohbuchi, Car-
olina Redondo-Cabrera, et al. SHREC’12 track: generic 3D
shape retrieval. In 5th Eurographics Conference on 3D Ob-
ject Retrieval, 2012. 3
[17] Bo Li, Yijuan Lu, Chunyuan Li, Afzal Godil, Tobias
Schreck, Masaki Aono, Qiang Chen, Nihad Karim Chowd-
hury, Bin Fang, Takahiko Furuya, et al. SHREC’14 track:
Large scale comprehensive 3D shape retrieval.
In Euro-
graphics Workshop on 3D Object Retrieval, 2014. 2
[18] Joerg Liebelt and Cordelia Schmid. Multi-view object class
detection with a 3D geometric model. In CVPR, pages 1688–
1695. IEEE, 2010. 2
[19] Tianqiang Liu, Siddhartha Chaudhuri, Vladimir G. Kim, Qi-
Xing Huang, Niloy J. Mitra, and Thomas Funkhouser. Cre-
ating consistent scene graphs using a probabilistic grammar.
ACM TOG, December 2014. 2
[20] Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice
Santorini. Building a large annotated corpus of english: The
Penn Treebank. Computational linguistics, 19(2):313–330,
1993. 1
[21] George A. Miller. WordNet: a lexical database for English.
CACM, 1995. 1, 2, 3, 4
[22] Niloy J Mitra, Mark Pauly, Michael Wand, and Duygu Cey-
lan. Symmetry in 3D geometry: Extraction and applications.
In Computer Graphics Forum, volume 32, pages 1–23, 2013.
7
[23] Fakir S. Nooruddin and Greg Turk. Simplification and repair
of polygonal models using volumetric techniques. Visualiza-
tion and Computer Graphics, IEEE Transactions on, 2003.
7
[24] Bryan C Russell and Antonio Torralba. Building a database
of 3D scenes from user annotations. In CVPR, 2009. 2
[25] Manolis Savva, Angel X. Chang, Gilbert Bernstein, Christo-
pher D. Manning, and Pat Hanrahan.
On being the right
scale: Sizing large collections of 3D models. In SIGGRAPH
Asia 2014 Workshop on Indoor Scene Understanding: Where
Graphics meets Vision, 2014. 7
[26] Manolis Savva,
Angel X. Chang,
and Pat Hanrahan.
Semantically-Enriched
3D
Models
for
Common-sense
Knowledge.
CVPR 2015 Workshop on Functionality,
Physics, Intentionality and Causality, 2015. 7
[27] Philip Shilane, Patrick Min, Michael Kazhdan, and Thomas
Funkhouser.
The Princeton shape benchmark.
In Shape
Modeling Applications. IEEE, 2004. 2, 3
[28] Shuran Song and Jianxiong Xiao. Sliding shapes for 3D ob-
ject detection in depth images. In ECCV, 2014. 1
[29] Atsushi Tatsuma, Hitoshi Koyanagi, and Masaki Aono. A
large-scale shape benchmark for 3D object retrieval: Toy-
ohashi shape benchmark. In Asia Pacific Signal and Infor-
mation Processing Association, 2012. 3
[30] Antonio Torralba, Bryan C Russell, and Jenny Yuen. La-
belMe: Online image annotation and applications. Proceed-
ings of the IEEE, 98(8):1467–1484, 2010. 7
[31] Remco C. Veltkamp and FB ter Harr. SHREC 2007 3D shape
retrieval contest. Technical report, Utrecht University Tech-
nical Report UU-CS-2007-015, 2007. 3
[32] Dejan V Vrani´c. 3D model retrieval. University of Leipzig,
Germany, PhD thesis, 2004. 3
10
[33] Raoul Wessel, Ina Bl¨umel, and Reinhard Klein. A 3D shape
benchmark for retrieval and automatic classification of ar-
chitectural data. In Eurographics 2009 Workshop on 3D Ob-
ject Retrieval, pages 53–56. The Eurographics Association,
2009. 3
[34] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Lin-
guang Zhang, Xiaoou Tang, and Jianxiong Xiao.
3D
ShapeNets: A Deep Representation for Volumetric Shapes.
CVPR, 2015. 1, 2, 4
[35] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond
PASCAL: A benchmark for 3D object detection in the wild.
In WACV, 2014. 2, 7
[36] Jianxiong Xiao, Andrew Owens, and Antonio Torralba.
SUN3D: A database of big spaces reconstructed using SfM
and object labels. In ICCV, pages 1625–1632, 2013. 2
[37] Juan Zhang, Kaleem Siddiqi, Diego Macrini, Ali Shokoufan-
deh, and Sven Dickinson. Retrieving articulated 3-D mod-
els using medial surfaces and their graph spectra. In Energy
minimization methods in computer vision and pattern recog-
nition, 2005. 3
A. Appendix
A.1. Hierarchical Rigid Alignment
In the following, we describe our hierarchical rigid align-
ment algorithm in more detail.
As a pre-processing step, we first semi-automatically
align the upright orientation of each shape.
Fortunately,
most shapes downloaded from the web are by default placed
in the upright orientations. For those that are not, we filter
them out by manual inspection. We then convert models to
point clouds through furthest point sampling and perform
PCA on the point sets. Finally, we ask a person to pick
the vector of correct upright orientation from six candidates
containing the PCA axes and their reverse directions.
Starting from a leaf category in ShapeNet, we jointly
align all shapes following prior work [8]. If a leaf cate-
gory has more than 100 shapes, we further partition it into
smaller, more coherent clusters by k-means clustering us-
ing pose-invariant global features, such as phase-invariant
HoG features [see appendix]. Here we briefly review [8].
Each shape is associated with a random variable, denot-
ing the transformation of the shape from its original pose
to the consistent canonical pose. Over the set of shapes, a
Markov Random Field (MRF) is constructed, whose energy
function measures the consistency of all pairs of shapes af-
ter applying their transformations. In practice, the space of
rigid transformations is discretized into N bins. We perform
MAP inference over the MRF to find the optimal transfor-
mation for each shape. We then manual inspect the results
and correct occasional errors.
After this step, we represent each leaf node category by
the shape in the centroid of the feature space. Then, we
gather the representative shapes for all leaf categories of an
intermediate category and apply [8] again for joint align-
ment. This higher-level algorithmic alignment is verified
by a person again. The procedure is applied along the tax-
onomy hierarchy until the root node is reached.
11