Structured3D: A Large Photo-realistic Dataset

for Structured 3D Modeling

Jia Zheng1,2∗† , Junfei Zhang1∗ , Jing Li2 , Rui Tang1 ,

Shenghua Gao2,3 , and Zihan Zhou4
1
KooLab, [Link] 2 ShanghaiTech University
3
Shanghai Engineering Research Center of Intelligent Vision and Imaging
arXiv:1908.00222v3 [[Link]] 17 Jul 2020

4
The Pennsylvania State University
[Link]

Abstract. Recently, there has been growing interest in developing

learning-based methods to detect and utilize salient semi-global or global
structures, such as junctions, lines, planes, cuboids, smooth surfaces, and
all types of symmetries, for 3D scene modeling and understanding. How-
ever, the ground truth annotations are often obtained via human labor,
which is particularly challenging and inefficient for such tasks due to the
large number of 3D structure instances (e.g., line segments) and other
factors such as viewpoints and occlusions. In this paper, we present a
new synthetic dataset, Structured3D, with the aim of providing large-
scale photo-realistic images with rich 3D structure annotations for a wide
spectrum of structured 3D modeling tasks. We take advantage of the
availability of professional interior designs and automatically extract 3D
structures from them. We generate high-quality images with an industry-
leading rendering engine. We use our synthetic dataset in combination
with real images to train deep networks for room layout estimation and
demonstrate improved performance on benchmark datasets.

Keywords: Dataset · 3D structure · Photo-realistic rendering

1 Introduction
Inferring 3D information from 2D sensory data such as images and videos has
long been a central research topic in computer vision. Conventional approach
to building 3D models typically relies on detecting, matching, and triangulating
local image features (e.g., patches, superpixels, edges, and SIFT features). Al-
though significant progress has been made over the past decades, these methods
still suffer from some fundamental problems. In particular, local feature detection
is sensitive to a large number of factors such as scene appearance (e.g., texture-
less areas and repetitive patterns), lighting conditions, and occlusions. Further,
the noisy, point cloud-based 3D model often fails to meet the increasing demand
for high-level 3D understanding in real-world applications.
*: Equal contribution.
†: The work was partially done when Jia Zheng interned at KooLab, [Link].
2 J. Zheng et al.

Floorplan Junctions and lines Planes Junction-line-plane relationship Original Lighting

3D design (a) Cuboid Manhattan world Semantics (b) Configuration Semantic labels (c)

Fig. 1: The Structured3D dataset. From a large collection of house designs (a)
created by professional designers, we automatically extract a variety of ground
truth 3D structure annotations (b) and generate photo-realistic 2D images (c).

When perceiving 3D scenes, humans are remarkably effective in using salient

global structures such as lines, contours, planes, smooth surfaces, symmetries,
and repetitive patterns. Thus, if a reconstruction algorithm can take advantage
of such global information, it is natural to expect the algorithm to obtain more
accurate results. Traditionally, however, it has been computationally challeng-
ing to reliably detect such global structures from noisy local image features.
Recently, deep learning-based methods have shown promising results in detect-
ing various forms of structure directly from the images, including lines [12,40],
planes [19,35,16,36], cuboids [10], floorplans [17,18], room layouts [14,41,26], ab-
stracted 3D shapes [28,32], and smooth surfaces [11].
With the fast development of deep learning methods comes the need for large
amounts of accurately annotated data. In order to train the proposed neural
networks, most prior work collects their own sets of images and manually label
the structure of interest in them. Such a strategy has several shortcomings. First,
due to the tedious process of manually labeling and verifying all the structure
instances (e.g., line segments) in each image, existing datasets typically have
limited sizes and scene diversity. And the annotations may also contain errors.
Second, since each study primarily focuses on one type of structure, none of these
datasets has multiple types of structure labeled. As a result, existing methods
are unable to exploit relations between different types of structure (e.g., lines
and planes) as humans do for effective, efficient, and robust 3D reconstruction.
In this paper, we present a large synthetic dataset with rich annotations
of 3D structure and photo-realistic 2D renderings of indoor man-made environ-
ments (Fig. 1). At the core of our dataset design is a unified representation of 3D
structure which enables us to efficiently capture multiple types of 3D structure
in the scene. Specifically, the proposed representation considers any structure
as relationship among geometric primitives. For example, a “wireframe” struc-
ture encodes the incidence and intersection relationship between line segments,
whereas a “cuboid” structure encodes the rotational and reflective symmetry
relationship among its planar faces. With our “primitive + relationship” repre-
sentation, one can easily derive the ground truth annotations for a wide variety of
semi-global and global structures (e.g., lines, wireframes, planes, regular shapes,
 Structured3D: A Large Photo-realistic Dataset for Structured 3D Modeling 3

Table 1: An overview of datasets with structure annotations. † : The actual num-

bers are not explicitly given and hard to estimate, because these datasets contain
images from Internet (LSUN Room Layout, PanoContext), or multiple sources
(LayoutNet). ∗ : Dataset is unavailable online at the time of publication.
Datasets #Scenes #Rooms #Frames Annotated structure
PlaneRCNN [16] - - 100,000 planes
Wireframe [12] - - 5,462 wireframe (2D)
SceneCity 3D [40] 230 - 23,000 wireframe (3D)
SUN Primitive [34] - - 785 cuboids, other primitives
LSUN Room Layout [39] - n/a† 5,394 cuboid layout
PanoContext [37] - n/a† 500 (pano) cuboid layout
LayoutNet [41] - n/a† 1,071 (pano) cuboid layout
MatterportLayout∗ [42] - n/a† 2,295 (RGB-D pano) Manhattan layout
Raster-to-Vector [17] 870 - - floorplan
Structured3D 3,500 21,835 196,515 “primitive + relationship”

floorplans, and room layouts), and also exploit their relations in future data-
driven approaches (e.g., the wireframe formed by intersecting planar surfaces in
the scene).
To create a large-scale dataset with the aim of facilitating research on data-
driven methods for structured 3D scene understanding, we leverage the avail-
ability of professional interior designs and millions of production-level 3D ob-
ject models – all coming with fine geometric details and high-resolution textures
(Fig. 1(a)). We first use computer programs to automatically extract information
about 3D structure from the original house design files. As shown in Fig. 1(b),
our dataset contains rich annotations of 3D room structure including a variety
of geometric primitives and relationships. To further generate photo-realistic 2D
images (Fig. 1(c)), we utilize industry-leading rendering engines to model the
lighting conditions. Currently, our dataset consists of more than 196k images of
21,835 rooms in 3,500 scenes (i.e., houses).
To showcase the usefulness and uniqueness of the proposed Structured3D
dataset, we train deep networks for room layout estimation on a subset of the
dataset. We show that the models trained on both synthetic and real data outper-
form the models trained on real data only. Further, following the spirit of [27,8],
we show how multi-modal annotations in our dataset can benefit domain adap-
tation tasks.
In summary, the main contributions of this paper are:
– We create the Structured3D dataset, which contains rich ground truth 3D
structure annotations of 21,835 rooms in 3,500 scenes, and more than 196k
photo-realistic 2D renderings of the rooms.
– We introduce a unified “primitive + relationship” representation. This rep-
resentation enables us to efficiently capture a wide variety of semi-global or
global 3D structures and their mutual relationships.
– We verify the usefulness of our dataset by using it to train deep networks for
room layout estimation and demonstrating improved performance on public
benchmarks.
4 J. Zheng et al.

(a) Plane [16] (b) Wireframe [12] (c) Cuboid [10] (d) Room layout [39]

(e) Floorplan [17] (f) Abstracted 3D shape (wireframe [32] and cuboid [28])

Fig. 2: Example annotations of structure in existing datasets. The reference num-

ber indicates the paper from which the illustration is originally from.

2 Related Work

Datasets. Table 1 summarizes existing datasets for structured 3D scene mod-

eling. Additionally, [28,32] provide datasets with structured representations of
single objects. We show example annotations in these datasets in Fig. 2. Note
that ground truth annotations in most datasets are manually labeled. This is
one main reason why all these datasets have limited size, i.e., contain no more
than a few thousand images. One exception is [16], which employs a multi-model
fitting algorithm to automatically extract planes from 3D scans in the ScanNet
dataset [9]. But such algorithms are sensitive to data noises and outliers, thus
introduce errors in the annotations (Fig. 2(a)). Similar to our work, SceneCity
3D [40] also contains synthetic images with ground truth automatically extracted
from CAD models. But the number of scenes is limited to 230. Further, none of
these datasets has more than one type of structure labeled, although different
types of structure often have strong relations among them. For example, from
the wireframe in Fig. 2(b) humans can easily identify other types of structure
such as planes and cuboids. Our new dataset sets to bridge the gap between
what is needed to train machine learning models to achieve human-level holistic
3D scene understanding and what is being offered by existing datasets.
Note that our dataset is very different from other popular large-scale 3D
datasets, such as NYU v2 [23], SUN RGB-D [24], 2D-3D-S [4,3], ScanNet [9],
and Matterport3D [6], in which the ground truth 3D information is stored in the
format of point clouds or meshes. These datasets lack ground truth annotations
of semi-global or global structures. While it is theoretically possible to extract
3D structure by applying structure detection algorithms to the point clouds or
meshes (e.g., extracting planes from ScanNet as did in [16]), the detection results
are often noisy and even contain errors. In addition, for some types of structure
 Structured3D: A Large Photo-realistic Dataset for Structured 3D Modeling 5

Table 2: Comparison of 3D scene datasets. † : Meshes are obtained by 3D recon-

struction algorithm. Notations for applications: O (object detection), U (scene
understanding), S (image synthesis), M (structured 3D modeling).
Datasets Scene design type 3D annotation 2D rendering Applications
NYU v2 [23] Real Raw RGB-D Real images OU
SUN RGB-D [24] Real Raw RGB-D Real images OU
2D-3D-S [4,3] Real Mesh† Real images OU
ScanNet [9] Real Mesh† Real images OU
Matterport3D [6] Real Mesh† Real images OU
SUNCG [25] Amateur Mesh n/a OU
SceneNet RGB-D [20] Random Mesh Photo-realistic OU
InteriorNet [15] Professional n/a Photo-realistic OUS
Structured3D Professional 3D structures Photo-realistic OUSM

like wireframes and room layouts, how to reliably detect them from raw sensor
data remains an active research topic in computer vision.
In recent years, synthetic datasets have played an important role in the suc-
cessful training of deep neural networks. Notable examples for indoor scene un-
derstanding include SUNCG [25], SceneNet RGB-D [20], and InteriorNet [15].
These datasets exceed real datasets in terms of scene diversity and frame num-
bers. But just like their real counterparts, these datasets lack ground truth
structure annotations. Another issue with some synthetic datasets is the de-
gree of realism in both the 3D models and the 2D renderings. [38] shows that
physically-based rendering could boost the performance of various indoor scene
understanding tasks. To ensure the quality of our dataset, we make use of 3D
room models created by professional designers and the state-of-the-art industrial
rendering engines. Table 2 summarizes the differences of 3D scene datasets.
Room layout estimation. Room layout estimation aims to reconstruct the
enclosing structure of the indoor scene, consisting of walls, floor, and ceiling.
Existing public datasets (e.g., PanoContext [37] and LayoutNet [41]) assume
a simple box-shaped layout. PanoContext [37] collects about 500 panoramas
from the SUN360 dataset [33], LayoutNet [41] extends the layout annotations
to include panoramas from 2D-3D-S [3]. Recently, MatterportLayout [42] col-
lects 2,295 RGB-D panoramas from Matterport3D [6] and extends annotations
to Manhattan layout. We note that all room layout in these real datasets is
manually labeled by the human. Since the room structure may be occluded by
furniture and other objects, the “ground truth” inferred by humans may not be
consistent with the actual layout. In our dataset, all ground truth 3D annotations
are automatically extracted from the original house design files.

3 A Unified Representation of 3D Structure

The main goal of our dataset is to provide rich annotations of ground truth 3D
structure. A naive way to do so is generating and storing different types of 3D an-
notations in the same format as existing works, like wireframes as in [12], planes
6 J. Zheng et al.

(a) Primitives: junctions and lines (b) Primitives: planes (c) Relationships: R1 and R2

(d) Relationships: R3 (e) Relationships: R4 (f) Relationships: R5

Fig. 3: The ground truth 3D structure annotations in our dataset are represented
by primitives and relationships. (a): Junctions and lines. (b): Planes. We high-
light the planes in a single room. (c): Plane-line and line-junction relationships.
We highlight a junction, the three lines intersecting at the junction, and the
planes intersecting at each of the lines. (d): Cuboids. We highlight one cuboid
instance. (e): Manhattan world. We use different colors to denote planes aligned
with different directions. (f ): Semantic objects. We highlight a “room”, a “bal-
cony”, and the “door” connecting them.

as in [16], floorplans as in [17], and so on. But this leads to a lot of redundancy.
For example, planes in man-made environments are often bounded by a number
of line segments, which are part of the wireframe. Even worse, by representing
wireframes and planes separately, the relationships between them are lost. In
this paper, we present a unified representation in order to minimize redundancy
while preserving mutual relationships. We show how the most common types
of structure studied in the literature (e.g., planes, cuboids, wireframes, room
layouts, and floorplans) can be derived from our representation.
Our representation of the structure is largely inspired by the early work of
Witkin and Tenenbaum [31], which characterizes structure as “a shape, pattern,
or configuration that replicates or continues with little or no change over an
interval of space and time”. Accordingly, to describe any structure, we need to
specify: (i) what pattern is continuing or replicating (e.g., a patch, an edge, or
a texture descriptor), and (ii) the domain of its replication or continuation. In
this paper, we call the former primitives and the latter relationships.

3.1 The “Primitive + Relationship” Representation

We now show how to describe a man-made environment using a unified represen-
tation. For ease of exposition, we assume all objects in the scene can be modeled
 Structured3D: A Large Photo-realistic Dataset for Structured 3D Modeling 7

by piece-wise planar surfaces. But our representation can be easily extended to

more general surfaces. An illustration of our representation is shown in Fig. 3.
Primitives. Generally, a man-made scene has the following geometric primi-
tives:
– Planes P: We model the scene as a collection of planes P = {p1 , p2 , . . .}.
Each plane is described by its parameters p = {n, d}, where n and d denote
the surface normal and the distance to the origin, respectively.
– Lines L: When two planes intersect in the 3D space, a line is created. We
use L = {l1 , l2 , . . .} to represent the set of all 3D lines in the scene.
– Junction points X: When two lines meet in the 3D space, a junction point
is formed. We use X = {x1 , x2 , . . .} to represent the set of all junction points.

Relationships. Next, we define some common types of relationships between

the geometric primitives:
– Plane-line relationships (R1 ): We use a matrix W1 to record all incidence
and intersection relationships between planes in P and lines in L. Specifically,
the ij-th entry of W1 is 1 if li is on pj , and 0 otherwise. Note that two planes
are intersected at some line if and only if the corresponding entry in W1T W1
is nonzero.
– Line-point relationships (R2 ): Similarly, we use a matrix W2 to record
all incidence and intersection relationships between lines in L and points
in X. Specifically, the mn-th entry of W2 is 1 if xm is on ln , and 0 other-
wise. Note that two lines are intersected at some junction if and only if the
corresponding entry in W2T W2 is nonzero.
– Cuboids (R3 ): A cuboid is a special arrangement of plane primitives with
rotational and reflection symmetry along x-, y- and z-axes. The correspond-
ing symmetry group is the dihedral group D2h .
– Manhattan world (R4 ): This is a special type of 3D structure commonly
used for indoor and outdoor scene modeling. It can be viewed as a grouping
relationship, in which all theSplane primitives can be grouped into three
3
classes, P1 , P2 , and P3 , P = i=1 Pi . Further, each class is represented by
a single normal vector ni , such that nTi nj = 0, i 6= j.
– Semantic objects (R5 ): Semantic information is critical for many 3D com-
puter vision tasks. It can be regarded as another type of grouping relation-
ship, in which each semantic object instance corresponds to one or more
primitives defined above. For example, each “wall”, “ceiling”, or “floor” in-
stance is associated with one plane primitive; each “chair” instance is asso-
ciated with a set of multiple plane primitives. Further, such a grouping is
hierarchical. For example, we can further group one floor, one ceiling, and
multiple walls to form a “living room” instance. And a “door” or a “window”
is an opening which connects two rooms (or one room and the outer space).
Note that the relationships are not mutually exclusive, in the sense that
a primitive can belong to multiple relationship instances of the same type or
different types. For example, a plane primitive can be shared by two cuboids,
8 J. Zheng et al.

and at the same time belong to one of the three classes in the Manhattan world
model.
Discussion. The primitives and relationships we discussed above are just a
few most common examples. They are by no means exhaustive. For example,
our representation can be easily extended to include other primitives such as
parametric surfaces. And besides cuboids, there are many other types of regular
or symmetric shapes in man-made environments, where type corresponds to a
different symmetry group.
Our representation of 3D structures is also related to the graph representa-
tions in semantic scene understanding [13,2,30]. As these graphs focus on seman-
tics, geometry is represented in simplified manners by (i) 6D object poses and
(ii) coarse, discrete spatial relations such as ”supported by”, ”front”, ”back”,
and ”adjacent”. In contrast, our representation focuses on modeling the scene
geometry using fine-grained primitives (i.e., junctions, lines, and planes) and
relationships (in terms of topology and regularities). Thus, it is highly com-
plementary to the scene graphs in prior work. Intuitively, it can be used for
geometric analysis and synthesis tasks, in a similar way as scene graphs are used
for semantic scene understanding.

3.2 Relation to Existing Models

Given our representation which contains primitives P = {P, L, X} and relation-

ships R = {R1 , R2 , . . .}, we show how several types of 3D structure commonly
studied in the literature can be derived from it. We again refer readers to Fig. 2
for illustrations of these structures.
Planes: A large volume of studies in the literature model the scene as a collection
of 3D planes, where each plane is represented by its parameters and boundary.
To generate such a model, we simply use the plane primitives P. For each p ∈ P,
we further obtain its boundary by using matrix W1 in R1 to find all the lines in
L that form an incidence relationship with p.
Wireframes: A wireframe consists of lines L and junction points P, and their
incidence and intersection relationships (R2 ).
Cuboids: This model is same as R3 .
Manhattan layouts: A Manhattan room layout model includes a “room” as
defined in R5 which also satisfies the Manhattan world assumption (R4 ).
Floorplans: A floorplan is a 2D vector representation that consists of a set of
line segments and semantic labels (e.g., room types). To obtain such a vector
representation, we can identify all lines in L and junction points in X which lie
on a “floor” (as defined in R5 ). To further obtain the semantic room labels, we
can project all “rooms”, “doors”, and “windows” (as defined in R5 ) to this floor.
Abstracted 3D shapes: In addition to room structures, our representation
can also be applied to individual 3D object models to create abstractions in the
form of wireframes or cuboids, as described above.
 Structured3D: A Large Photo-realistic Dataset for Structured 3D Modeling 9

(a) (b)

Fig. 4: Comparison of 3D house designs. (a): The 3D models in our database are
created by professional designers using high-quality furniture models from world-
leading manufacturers. Most designs are being used in real-world production.
(b): The 3D models in SUNCG dataset [25] are created using Planner 5D [1],
an online tool for amateur interior design.

4 The Structured3D Dataset

Our unified representation enables us to encode a rich set of geometric prim-
itives and relationships for structured 3D modeling. With this representation,
our ultimate goal is to build a dataset that can be used to train machines to
achieve the human-level understanding of the 3D environment.
As a first step towards this goal, in this section, we describe our ongoing
effort to create a large-scale dataset of indoor scenes which include (i) ground
truth 3D structure annotations of the scene and (ii) realistic 2D renderings of the
scene. Note that in this work we focus on extracting ground truth annotations on
the room structure only. We plan to extend our dataset to include 3D structure
annotations of individual furniture models in the future.
In the following, we describe our general procedure to create the dataset.
We refer readers to the supplementary materials for additional details, including
dataset statistics and example annotations.

4.1 Extraction of Structured 3D Models

To extract a “primitive + relationship” scene representation, we utilize a large
database of house designs hand-crafted by professional designers. An example
design is shown in Fig. 4(a). All information of the design is stored in an industry-
standard format in the database so that specifications about the geometry (e.g.,
the precise size of each wall), textures and materials, and functions (e.g., which
room the wall belongs to) of all objects can be easily retrieved.
From the database, we have selected 3,500 house designs with 21,835 rooms.
We created a computer program to automatically extract all the geometric
primitives associated with the room structure, which consists of the ceiling,
floor, walls, and openings (doors and windows). Given the precise measurements
and associated information of these entities, it is straightforward to generate all
planes, lines, and junctions, as well as their relationships (R1 and R2 ).
10 J. Zheng et al.

(a) Original (b) Simple configuration (c) Empty configuration

(d) Lighting (e) Depth (f) Semantic labels

Fig. 5: Examples of our rendered panoramic images.

Since the measurements are highly accurate and noise-free, other types of
relationship such a Manhattan world (R3 ) and cuboids (R4 ) can also be eas-
ily obtained by clustering the primitives, followed by a geometric verification
process. Finally, to include semantic information (R5 ) into our representation,
we map the relevant labels provided by the professional designers to the geo-
metric primitives in our representation. Fig. 3 shows examples of the extracted
geometric primitives and relationships.

4.2 Photo-realistic 2D Rendering

To ensure the quality of our 2D renderings, our rendering engine is developed

in collaboration with a company specialized in interior design rendering. Our
engine uses a well-known ray-tracing method [21], a Monte Carlo approach to
approximating realistic Global Illumination (GI), for RGB rendering. The other
ground truth images are obtained by a customized path-tracer renderer on top of
Intel Embree [29], an open-source collection of ray-tracing kernels for x86 CPUs.
Each room is manually created by professional designers with over one mil-
lion CAD models of furniture from world-leading manufacturers. These high-
resolution furniture models are measured in real-world dimensions and being
used in real production. A default lighting setup is also provided. Fig. 4 com-
pares the 3D models in our database with those in SUNCG [25], which are
created using Planner 5D [1], an online tool for amateur interior design.
At the time of rendering, a panoramic or pin-hole camera is placed at random
locations not occupied by objects in the room. We use 512 × 1024 resolution for
panoramas and 720 × 1280 for perspective images. Fig. 5 shows example panora-
mas rendered by our engine. For each room, we generate different configurations
(full, simple, and empty) by removing some or all the furniture. We also mod-
ify the lighting setup to generate images with different temperatures. For each
image, our dataset also includes the depth map and semantic mask. Fig. 6 illus-
trates the degree of photo-realism of our dataset, where we compare the rendered
images with photos of real decoration guided by the design.
 Structured3D: A Large Photo-realistic Dataset for Structured 3D Modeling 11

Fig. 6: Photo-realistic rendering vs. real-world decoration. The first and third
columns are rendered images.

4.3 Use Cases

Due to the unique characteristics of our dataset, we envision it contributing to
computer vision research in terms of both methodology and applications.
Methodology. As our dataset contains multiple types of 3D structure anno-
tations as well as ground truth labels (e.g., semantic maps, depth maps, and
3D object bounding boxes), it enables researchers to design novel multi-modal
or multi-task approaches for a variety of vision tasks. As an example, we show
in Section 5 that, by leveraging multi-modal annotations in our dataset, we can
boost the performance of existing room layout estimation methods in the domain
adaptation framework.
Applications. Our dataset also facilitates research on a number of problems and
applications. For example, as shown in Table 1, all publicly available datasets for
room layout estimation are limited to simple cuboid rooms. Our dataset is the
first to provide the general (non-cuboid) room layout annotations. As another
example, existing datasets for floorplan reconstruction [18,7] contain about 100-
150 scenes, whereas our dataset includes 3,500 scenes.
Another major line of research that would benefit from our dataset is image
synthesis. With a photo-realistic rendering engine, we are able to generate images
given any scene configurations and viewpoints. These images may be used as
ground truth for tasks including image inpainting (e.g., completing an image
when certain furniture is removed) and novel view synthesis.
Finally, we would like to emphasize the potential of our dataset in terms
of extension capabilities. As we mentioned before, the unified representation
enables us to include many other types of structure in the dataset. As for 2D
rendering, depending on the application, we can easily simulate different effects
such as lighting conditions, fisheye and novel camera designs, motion blur, and
imaging noise. Furthermore, the dataset may be extended to include videos for
applications such as visual SLAM [5].

5 Experiments
5.1 Experiment Setup
To demonstrate the benefits of our dataset, we use it to train deep neural net-
works for room layout estimation, an important task in structured 3D modeling.
12 J. Zheng et al.

Table 3: Room layout statistics. † : MatterportLayout is the only other dataset

with non-cuboid layout annotations, but is unavailable at the time of publication.
#Corners 4 5 6 7 8 9 10+ Total
MatterportLayout† 1211 0 501 0 309 0 274 2295
Structured3D 13743 52 3727 30 1575 17 2691 21835

Real dataset. We use the same dataset as LayoutNet [41]. The dataset consists
of images from PanoContext [37] and 2D-3D-S [3], including 818 training images,
79 validation images, and 166 test images. Note that both datasets only provide
cuboid layout annotations.
Our Structured3D dataset. In this experiment, we use a subset of panoramas
with the original lighting and full configuration. Each panorama corresponds to
a different room in our dataset. We show statistics of different room layouts in
our dataset in Table 3. Since the current real dataset only contains cuboid layout
annotations (i.e., 4 corners), we choose 12k panoramic images with the cuboid
layout in our dataset. We split the images into 10k for training, 1k for validation,
and 1k for testing.
Evaluation metrics. Following [41,26], we adopt three standard metrics: (i)
3D IoU: intersection over union between predicted 3D layout and the ground
truth, (ii) Corner Error (CE): normalized `2 distance between predicted corner
and ground truth, and (iii) Pixel Error (PE): pixel-wise error between predicted
plane classes and ground truth.
Baselines. We choose two recent CNN-based approaches, LayoutNet [41,42]1
and HorizonNet [26]2 , based on their performance and source code availability.
LayoutNet uses a CNN to predict a corner probability map and a boundary map
from the panorama and vanishing lines, then optimizes the layout parameters
based on network predictions. HorizonNet represents room layout as three 1D
vectors, i.e., boundary positions of floor-wall, and ceiling-wall, and the existence
of wall-wall boundary. It trains CNNs to directly predict the three 1D vectors.
In this paper, we follow the default training setting of the respective methods.
For specific training procedures, please refer to the supplementary materials.

5.2 Experiment Results

Augmenting real datasets. In this experiment, we train LayoutNet and Hori-

zonNet in four different manners: (i) training only on our synthetic dataset (“s”),
(ii) training only on the real dataset (“r”), (iii) training on the synthetic and
real dataset with Balanced Gradient Contribution (BGC) [22] (“s + r”), and
(iv) pre-training on our synthetic dataset, then fine-tuning on the real dataset
(“s → r”). We adopt the training set of LayoutNet as the real dataset in this
experiment. The results are shown in Table 4. As one can see, augmenting real
1
[Link]
2
[Link]
 Structured3D: A Large Photo-realistic Dataset for Structured 3D Modeling 13

Table 4: Quantitative evaluation under different training schemes. The best and
the second best results are boldfaced and underlined, respectively.
PanoContext 2D-3D-S
Methods Config.
3D IoU (%) ↑ CE (%) ↓ PE (%) ↓ 3D IoU (%) ↑ CE (%) ↓ PE (%) ↓
s 75.64 1.31 4.10 57.18 2.28 7.55
r 84.15 0.64 1.80 83.39 0.74 2.39
LayoutNet [41,42]
s+r 84.96 0.61 1.75 83.66 0.71 2.31
s→r 84.77 0.63 1.89 84.04 0.66 2.08
s 75.89 1.13 3.15 67.66 1.18 3.94
r 83.42 0.73 2.09 84.33 0.64 2.04
HorizonNet [26]
s+r 84.45 0.70 1.89 84.36 0.59 1.90
s→r 85.27 0.66 1.86 86.01 0.61 1.84

Table 5: Quantitative evaluation using varying synthetic data size in pre-training.

The best and the second best results are boldfaced and underlined, respectively.
Synthetic PanoContext 2D-3D-S
Methods
Data Size 3D IoU (%) ↑ CE (%) ↓ PE (%) ↓ 3D IoU (%) ↑ CE (%) ↓ PE (%) ↓
1k 83.81 0.66 1.99 83.57 0.72 2.31
LayoutNet [41,42] 5k 84.47 0.67 1.97 84.55 0.69 2.21
10k 84.77 0.63 1.89 84.04 0.66 2.08
1k 83.77 0.74 2.11 85.19 0.63 2.01
HorizonNet [26] 5k 84.13 0.73 2.07 86.35 0.61 1.87
10k 85.27 0.66 1.86 86.01 0.61 1.84

datasets with our synthetic data boosts the performance of both networks. We
refer readers to supplementary materials for more qualitative results.
Performance vs. synthetic data size. We further study the relationship
between the number of synthetic images used in pre-training and the accuracy
on the real dataset. We sample 1k, 5k and 10k synthetic images for pre-training,
then fine-tune the model on the real dataset. The results are shown in Table 5.
As expected, using more synthetic data generally improves the performance.
Domain adaptation. Domain adaptation techniques (e.g., [27]) have been
shown to be effective in bridging the performance gap when directly applying
models learned on synthetic data to real environments. In this experiment, we
do not assume access to ground truth layout labels in the real dataset. We adopt
LayoutNet as the task network and use PanoContext and 2D-3D-S separately.
We apply a discriminator network to align the output features of the LayoutNet
for two domains. Inspired by [8], we further leverage multi-modal annotations
in our dataset by adding another decoder branch to the LayoutNet for depth
prediction. We concatenate the boundary, corner, and depth predictions as the
input of the discriminator network. The results are shown in the Table 6. By
incorporating additional information, i.e., depth map, we further boost the per-
formance on both datasets. This illustrates the advantage of including multiple
types of ground truth in our dataset.
Limitation of real datasets. Due to human errors, the annotation in real
datasets is not always consistent with the actual room layout. In the left image
14 J. Zheng et al.

Table 6: Domain adaptation results. NA: non-adaptive baseline. +DA: align lay-
out estimation output. +Depth: align both layout estimation and depth outputs.
Real: train in the target domain.
PanoContext 2D-3D-S
Methods
3D IoU (%) ↑ CE (%) ↓ PE (%) ↓ 3D IoU (%) ↑ CE (%) ↓ PE (%) ↓
NA 75.64 1.31 4.10 57.18 2.28 7.55
+DA 76.91 1.19 3.64 70.08 1.36 4.66
+Depth 78.34 1.03 2.99 72.99 1.24 3.60
Real 81.76 0.95 2.58 81.82 0.96 3.13

Fig. 7: Limitation of real datasets. Left: PanoContext dataset. Right: 2D-3D-S

dataset. Blue lines are ground truth layout and green lines are predictions.

of Fig. 7, the room is a non-cuboid layout, but the ground truth layout is labeled
as cuboid shape. In the right image, the front wall is not labeled as ground truth.
These examples illustrate the limitation of using real datasets as benchmarks.
We avoid such errors in our dataset by automatically generating ground truth
from the original design files.

6 Conclusion
In this paper, we present Structured3D, a large synthetic dataset with rich
ground truth 3D structure annotations of 21,835 rooms and more than 196k
photo-realistic 2D renderings. Among many potential use cases of our dataset,
we further demonstrate its benefit in augmenting real data and facilitating do-
main adaptation for the room layout estimation task.
We view this work as an important and exciting step towards building intel-
ligent machines which can achieve human-level holistic 3D scene understanding.
In the future, we will continue to add more 3D structure annotations of the
scenes and objects to the dataset, and explore novel ways to use the dataset to
advance techniques for structured 3D modeling and understanding.
Acknowledgement. We would like to thank [Link] for providing the
database of house designs and the rendering engine. We especially thank Qing
Ye and Qi Wu from [Link] for the help on the data rendering. This
work was partially supported by the National Key R&D Program of China
(#2018AAA0100704) and the National Science Foundation of China (#61932020).
Zihan Zhou was supported by NSF award #1815491.
 Structured3D: A Large Photo-realistic Dataset for Structured 3D Modeling 15

References

1. Planner 5d. [Link] 9, 10

2. Armeni, I., He, Z.Y., Gwak, J., Zamir, A.R., Fischer, M., Malik, J., Savarese, S.:
3d scene graph: A structure for unified semantics, 3d space, and camera. In: ICCV.
pp. 5664–5673 (2019) 8
3. Armeni, I., Sax, A., Zamir, A.R., Savarese, S.: Joint 2d-3d-semantic data for indoor
scene understanding. CoRR abs/1702.01105 (2017) 4, 5, 12
4. Armeni, I., Sener, O., Zamir, A.R., Jiang, H., Brilakis, I.K., Fischer, M., Savarese,
S.: 3d semantic parsing of large-scale indoor spaces. In: CVPR. pp. 1534–1543
(2016) 4, 5
5. Cadena, C., Carlone, L., Carrillo, H., Latif, Y., Scaramuzza, D., Neira, J., Reid,
I.D., Leonard, J.J.: Past, present, and future of simultaneous localization and map-
ping: Toward the robust-perception age. IEEE Trans. Robotics 32(6), 1309–1332
(2016) 11
6. Chang, A.X., Dai, A., Funkhouser, T.A., Halber, M., Nießner, M., Savva, M.,
Song, S., Zeng, A., Zhang, Y.: Matterport3d: Learning from RGB-D data in indoor
environments. In: 3DV. pp. 667–676 (2017) 4, 5
7. Chen, J., Liu, C., Wu, J., Furukawa, Y.: Floor-sp: Inverse CAD for floorplans by
sequential room-wise shortest path. In: ICCV. pp. 2661–2670 (2019) 11
8. Chen, Y., Li, W., Chen, X., Van Gool, L.: Learning semantic segmentation from
synthetic data: A geometrically guided input-output adaptation approach. In:
CVPR. pp. 1841–1850 (2019) 3, 13
9. Dai, A., Chang, A.X., Savva, M., Halber, M., Funkhouser, T., Nießner, M.: Scannet:
Richly-annotated 3d reconstructions of indoor scenes. In: CVPR. pp. 5828–5839
(2017) 4, 5
10. Dwibedi, D., Malisiewicz, T., Badrinarayanan, V., Rabinovich, A.: Deep cuboid
detection: Beyond 2d bounding boxes. CoRR abs/1611.10010 (2016) 2, 4
11. Groueix, T., Fisher, M., Kim, V.G., Russell, B., Aubry, M.: A papier-mâché ap-
proach to learning 3d surface generation. In: CVPR. pp. 216–224 (2018) 2
12. Huang, K., Wang, Y., Zhou, Z., Ding, T., Gao, S., Ma, Y.: Learning to parse
wireframes in images of man-made environments. In: CVPR. pp. 626–635 (2018)
2, 3, 4, 5
13. Huang, S., Qi, S., Zhu, Y., Xiao, Y., Xu, Y., Zhu, S.C.: Holistic 3d scene parsing
and reconstruction from a single rgb image. In: ECCV. pp. 194–211 (2018) 8
14. Lee, C., Badrinarayanan, V., Malisiewicz, T., Rabinovich, A.: Roomnet: End-to-
end room layout estimation. In: ICCV. pp. 4875–4884 (2017) 2
15. Li, W., Saeedi, S., McCormac, J., Clark, R., Tzoumanikas, D., Ye, Q., Huang,
Y., Tang, R., Leutenegger, S.: Interiornet: Mega-scale multi-sensor photo-realistic
indoor scenes dataset. In: BMVC. p. 77 (2018) 5
16. Liu, C., Kim, K., Gu, J., Furukawa, Y., Kautz, J.: Planercnn: 3d plane detection
and reconstruction from a single image. In: CVPR. pp. 4450–4459 (2019) 2, 3, 4,
6
17. Liu, C., Wu, J., Kohli, P., Furukawa, Y.: Raster-to-vector: Revisiting floorplan
transformation. In: ICCV. pp. 2214–2222 (2017) 2, 3, 4, 6
18. Liu, C., Wu, J., Furukawa, Y.: Floornet: A unified framework for floorplan recon-
struction from 3d scans. In: ECCV. pp. 203–219 (2018) 2, 11
19. Liu, C., Yang, J., Ceylan, D., Yumer, E., Furukawa, Y.: Planenet: Piece-wise planar
reconstruction from a single rgb image. In: CVPR. pp. 2579–2588 (2018) 2
16 J. Zheng et al.

20. McCormac, J., Handa, A., Leutenegger, S., Davison, A.J.: Scenenet RGB-D: can
5m synthetic images beat generic imagenet pre-training on indoor segmentation?
In: ICCV. pp. 2697–2706 (2017) 5
21. Purcell, T.J., Buck, I., Mark, W.R., Hanrahan, P.: Ray tracing on programmable
graphics hardware. ACM Trans. Graph. 21(3), 703–712 (2002) 10
22. Ros, G., Stent, S., Alcantarilla, P.F., Watanabe, T.: Training constrained deconvo-
lutional networks for road scene semantic segmentation. CoRR abs/1604.01545
(2016) 12
23. Silberman, N., Hoiem, D., Kohli, P., Fergus, R.: Indoor segmentation and support
inference from RGBD images. In: ECCV. pp. 746–760 (2012) 4, 5
24. Song, S., Lichtenberg, S.P., Xiao, J.: SUN RGB-D: A RGB-D scene understanding
benchmark suite. In: CVPR. pp. 567–576 (2015) 4, 5
25. Song, S., Yu, F., Zeng, A., Chang, A.X., Savva, M., Funkhouser, T.A.: Semantic
scene completion from a single depth image. In: CVPR. pp. 1746–1754 (2017) 5,
9, 10
26. Sun, C., Hsiao, C.W., Sun, M., Chen, H.T.: Horizonnet: Learning room layout with
1d representation and pano stretch data augmentation. In: CVPR. pp. 1047–1056
(2019) 2, 12, 13
27. Tsai, Y.H., Hung, W.C., Schulter, S., Sohn, K., Yang, M.H., Chandraker, M.:
Learning to adapt structured output space for semantic segmentation. In: CVPR.
pp. 7472–7481 (2018) 3, 13
28. Tulsiani, S., Su, H., Guibas, L.J., Efros, A.A., Malik, J.: Learning shape abstrac-
tions by assembling volumetric primitives. In: CVPR. pp. 2635–2643 (2017) 2,
4
29. Wald, I., Woop, S., Benthin, C., Johnson, G.S., Ernst, M.: Embree: a kernel frame-
work for efficient CPU ray tracing. ACM Trans. Graph. 33(4), 143:1–143:8 (2014)
10
30. Wang, K., Lin, Y.A., Weissmann, B., Savva, M., Chang, A.X., Ritchie, D.: Planit:
Planning and instantiating indoor scenes with relation graph and spatial prior
networks. ACM Trans. Graph. 38(4) (2019) 8
31. Witkin, A.P., Tenenbaum, J.M.: On the role of structure in vision. In: Beck, J.,
Hope, B., Rosenfeld, A. (eds.) Human and Machine Vision, pp. 481–543. Academic
Press (1983) 6
32. Wu, J., Xue, T., Lim, J.J., Tian, Y., Tenenbaum, J.B., Torralba, A., Freeman,
W.T.: 3d interpreter networks for viewer-centered wireframe modeling. IJCV
126(9), 1009–1026 (2018) 2, 4
33. Xiao, J., Ehinger, K.A., Oliva, A., Torralba, A.: Recognizing scene viewpoint using
panoramic place representation. In: CVPR. pp. 2695–2702 (2012) 5
34. Xiao, J., Russell, B., Torralba, A.: Localizing 3d cuboids in single-view images. In:
NeurIPS. pp. 746–754 (2012) 3
35. Yang, F., Zhou, Z.: Recovering 3d planes from a single image via convolutional
neural networks. In: ECCV. pp. 87–103 (2018) 2
36. Yu, Z., Zheng, J., Lian, D., Zhou, Z., Gao, S.: Single-image piece-wise planar 3d
reconstruction via associative embedding. In: CVPR. pp. 1029–1037 (2019) 2
37. Zhang, Y., Song, S., Tan, P., Xiao, J.: Panocontext: A whole-room 3d context
model for panoramic scene understanding. In: ECCV. pp. 668–686 (2014) 3, 5, 12
38. Zhang, Y., Song, S., Yumer, E., Savva, M., Lee, J.Y., Jin, H., Funkhouser, T.:
Physically-based rendering for indoor scene understanding using convolutional neu-
ral networks. In: CVPR. pp. 5287–5295 (2017) 5
39. Zhang, Y., Yu, F., Song, S., Xu, P., Seff, A., Xiao, J.: Large-scale scene under-
standing challenge: Room layout estimation (2016) 3, 4
 Structured3D: A Large Photo-realistic Dataset for Structured 3D Modeling 17

40. Zhou, Y., Qi, H., Zhai, S., Sun, Q., Chen, Z., Wei, L.Y., Ma, Y.: Learning to
reconstruct 3d manhattan wireframes from a single image. In: ICCV. pp. 7698–
7707 (2019) 2, 3, 4
41. Zou, C., Colburn, A., Shan, Q., Hoiem, D.: Layoutnet: Reconstructing the 3d room
layout from a single RGB image. In: CVPR. pp. 2051–2059 (2018) 2, 3, 5, 12, 13
42. Zou, C., Su, J., Peng, C., Colburn, A., Shan, Q., Wonka, P., Chu, H., Hoiem,
D.: 3d manhattan room layout reconstruction from a single 360 image. CoRR
abs/1910.04099 (2019) 3, 5, 12, 13