Parsing IKEA Objects: Fine Pose Estimation
Joseph J. Lim Hamed Pirsiavash Antonio Torralba
Massachusetts Institute of Technology
Compter Science and Artificial Intelligence Laboratory
{lim,hpirsiav,torralba}@csail.mit.edu
Abstract
We address the problem of localizing and estimating the
fine-pose of objects in the image with exact 3D models. Our
main focus is to unify contributions from the 1970s with re-
cent advances in object detection: use local keypoint de-
tectors to find candidate poses and score global alignment
of each candidate pose to the image. Moreover, we also
provide a new dataset containing fine-aligned objects with
their exactly matched 3D models, and a set of models for
widely used objects. We also evaluate our algorithm both
on object detection and fine pose estimation, and show that
our method outperforms state-of-the art algorithms.
1. Introduction
Just as a thought experiment imagine that we want to
detect and fit 3D models of IKEA furniture in the images
as shown in Figure 1. We can find surprisingly accurate
3D models of IKEA furniture, such as billy bookcase and
ektorp sofa, created by IKEA fans from Google 3D Ware-
house and other publicly available databases. Therefore,
detecting those models in images could seem to be a very
similar task to an instance detection problem in which we
have training images of the exact instance that we want to
detect. But, it is not exactly like detecting instances. In the
case of typical 3D models (including IKEA models from
Google Warehouse), the exact appearance of each piece is
not available, only the 3D shape is. Moreover, appearance
in real images will vary significantly due to a number of
factors. For instance, IKEA furniture might appear with
different colors and textures, and with geometric deforma-
tions (as people building them might not do it perfectly) and
occlusions (e.g., a chair might have a cushion on top of it).
The problem that we introduce in this paper is detect-
ing and accurately fitting exact 3D models of objects to real
images, as shown in Figure 1. Detecting 3D objects in im-
ages and estimating their pose was a popular topic in the
early days of computer vision [14] and has gained a renewed
interest in the last few years. The traditional approaches
3D Model Original Image
Fine-pose Estimation
Figure 1. Our goal in this paper is to detect and estimate the fine-
pose of an object in the image given an exact 3D model.
[12, 14] were dominated by using accurate geometric rep-
resentations of 3D objects with an emphasis on viewpoint
invariance. Objects would appear in almost any pose and
orientation in the image. Most of the approaches were de-
signed to work at instance level detection as it was expected
that the 3D model would accurately fit the model in the im-
age. Instance level detection regained interest with the in-
troduction of new invariant local descriptors that dramati-
cally improved the detection of interest points [17]. Those
models assumed knowledge about geometry and appear-
ance of the instance. Having access to accurate knowledge
about those two aspects allowed precise detections and pose
estimations of the object on images even in the presence of
clutter and occlusions.
In the last few years, researchers interested in category
1
Edgemap HOG Local correspondence deteciton Image
Figure 2. Local correspondence: for each 3D interest point X
i
(red, green, and blue), we train an LDA patch detector on an edgemap
and use its response as part of our cost function. We compute HOG on edgemaps to ensure a real image and our model share the modality.
level detection have extended 2D constellation models to in-
clude 3D information in the object representation. Many of
these models [9, 20, 10, 6, 5, 19, 16, 7, 21] rely on gradient-
based features [1, 13]. Category level detection requires the
models to be generic and flexible, as they have to deal with
all the variations in shape and appearance of the instances
that belong to the same category. Therefore, the shape rep-
resentations used for those models are coarse (e.g., modeled
as the constellation of a few 3D parts or planes) and the ex-
pected output is at best an approximate fitting of the 3D
model to the image.
In this paper we introduce a detection task that is in the
intersection of these two settings; it is more generic than
detecting instances, but we assume richer models than the
ones typically used in category level detection. In particu-
lar, we assume that accurate CAD models of the objects are
available. Hence, there is little variation on the shape of the
instances that form one category. However, there might be
large variation in the appearance, as the CAD models do not
completely constrain the appearance of the objects placed
in the real world. Although assuming that CAD models are
available might seem to be a restrictive assumption, there
are available 3D CAD models for most man-made artifacts,
used for manufacturing or virtual reality. All those mod-
els could be used as training data for our system. Hence,
we focus on detection and pose estimation of objects in the
wild given their 3D CAD models. Our goal is to provide an
accurate localization of the object, as in the instance level
detection problem, but dealing with some of the variability
that one finds in category level detection.
Our contributions are three folds: (1) Proposing a detec-
tion problem that has some of the challenges of category
level detection while allowing for an accurate representa-
tion of the object pose in the image. This problem will mo-
tivate the development of better 3D object models and the
algorithms needed to find them in images. (2) We propose
a novel solution that unifies contributions from the 1970s
with recent advances in object detection. (3) And we in-
troduce a new dataset of 3D IKEA models obtained from
Google Warehouse and real images containing instances of
IKEA furniture and annotated with ground truth pose.
2. Methods
We now propose the framework that can detect objects
and estimate their poses simultaneously by matching to one
of the 3D models in our database.
2.1. Our Model
The core of our algorithm is to define the final score
based on multiple sources of information such as local
correspondence, geometric distance, and global alignment.
Suppose we are given the image I containing an object for
which we want to estimate the projection matrix P with 9
degrees of freedom
1
. We define our cost function S with
three terms as follows:
S(P, c)=L(P, c)+w
D
D(P, c)+w
G
G(P ) (1)
where c refers to the set of correspondences, L measures er-
ror in local correspondences between the 3D model and 2D
image, D measures geometric distance between correspon-
dences in 3D, and G measures the global dissimilarity in
2D. Note that we designed our model to be linear in weight
vectors of w
D
and w
G
. We will use this linearity later to
learn our discriminative classifier.
2.2. Local correspondence error
The goal here is to measure the local correspondences.
Given a projection P , we find the local shape-based match-
ing score between the rendered image of the CAD model
and the 2D image. Because our CAD model contains only
1
We assume the image is captured by a regular pinhole camera and is
not cropped, so the the principal point is in the middle of image.
shape information, a simple distance measure on standard
local descriptors fails to match the two images. In order
to overcome this modality difference, we compute HOG on
the edgemap of both images since it is more robust to ap-
pearance change but still sensitive to the change in shape.
Moreover, we see better performance by training an LDA
classifier to discriminate each keypoint from the rest of the
points. This is equivalent to a dot product between descrip-
tors in the whitened space. One advantage of using LDA
is its high training speed compared to other previous HOG-
template based approaches. Figure 2 illustrates this step.
More formally, our correspondence error is measured by
L(P, c)=
X
i
H(
T
i
(x
c
i
) ↵ ) (2)
i
=⌃
1
((P (X
i
)) µ) (3)
where
i
is the weight learned using LDA based on the
covariance ⌃ and mean µ obtained from a large external
dataset [8], (·) is a HOG computed on a 20⇥20 pixel
edgemap patch of a given point, x
c
i
is the 2D correspond-
ing point of 3D point X
i
, P (·) projects a 3D point to 2D
coordinate given pose P , and lastly H(x ↵) binarizes x
to 0 if x ↵ or to 1 otherwise.
2.3. Geometric distance between correspondences
Given a proposed set of correspondences c between the
CAD model and the image, it is also necessary to measure if
c is a geometrically acceptable pose based on the 3D model.
For the error measure, we use euclidean distance between
the projection of x
i
and its corresponding 2D point x
c
i
; as
well as the line distance defined in [12] between the 3D line
l and its corresponding 2D line c
l
.
D(P,c)=
X
i2V
kP (X
i
) x
c
i
k
2
+ (4)
X
l2L
cos
c
l
, sin
c
l
, ⇢
c
l
P (X
l
1
) P (X
l
2
)
11
2
where X
l
1
and X
l
2
are end points of line l , and
c
l
and ⇢
c
l
are polar coordinate parameters of line c
l
.
The first term measures a pairwise distance between a 3D
interest point X
i
and its 2D corresponding point x
c
i
. The
second term measures the alignment error between a 3D in-
terest line l and one of its corresponding 2D lines c
l
[12].
We use the Hough transform on edges to extract 2D lines
from I. Note that this cost does not penalize displacement
along the line. This is important because the end points
of detected lines on I are not reliable due to occlusion and
noise in the image.
2.4. Global dissimilarity
One key improvement of our work compared to tradi-
tional works on pose estimation using a 3D model is how
we measure the global alignment. We use recently devel-
oped features to measure the global alignment between our
proposal and the image:
G(P )=[f
HOG
,f
region
,f
edge
,f
corr
,f
texture
], (5)
The description of each feature follows:
HOG-based: While D and L from Eq 1 are designed to
capture fine local alignments, the local points are sparse and
hence it is yet missing an object-scale alignment. In order
to capture edge alignment per orientation, we compute a
fine HOG descriptor (2x2 per cell) of edgemaps of I and
the rendered image of pose P . We use a similarity measure
based on cosine similarity between vectors .
f
HOG
=
(I)
T
(P )
k(P )k
2
,
(I)
T
(P )
k(I)M k
2
,
(I)
T
(P )
k(I)M kk(P )k
, (6)
where (·) is a HOG descriptor and M is a mask matrix
for counting how many pixels of P fall into each cell. We
multiply (I) by M to normalize only within the proposed
area (by P ) without being affected by other area.
Regions: We add another alignment feature based on super-
pixels. If our proposal P is a reasonable candidate, super-
pixels should not cross over the object boundary, except in
heavily occluded areas. To measure this spill over, we first
extract super-pixels, R
I
, from I using [3] and compute a
ratio between an area of a proposed pose, |R
P
|, and an area
of regions that has some overlap with R
P
. This ratio will
control the spillover effect as shown in Figure 3.
(a) Original image (b) Candidate 1 (c) Candidate 2
Figure 3. Region feature: One feature to measure a fine align-
ment is the ratio between areas of a proposed pose (yellow) and
regions overlapped with the proposed pose. (a) is an original im-
age I, and (b) shows an example to encourage, while (c) shows an
example to penalize.
f
region
=
"
P
|R
P
\R
I
|>0.1|R
P
|
|R
P
|
|R
I
|
#
(7)
Texture boundary: The goal of this feature is to capture
appearance by measuring how well our proposed pose sep-
arates object boundary. In other words, we would like to
measure the alignment between the boundary of our pro-
posed pose P and the texture boundary of I. For this pur-
pose, we use a well-known texture classification feature,
Local Binary Pattern (LBP) [15].
We compute histograms of LBP on inner boundary and
outer boundary of proposed pose P . We define an inner
boundary region by extruding the proposed object mask
followed by subtracting the mask, and an outer boundary
by diluting the proposed object mask. Essentially, the his-
tograms will encode the texture patterns of near-inner/outer
pixels along the proposed pose P ’s boundary. Hence, a
large change in these two histograms indicates a large
texture pattern change, and it is ideal if the LBP histogram
difference along the proposed pose’s boundary is large.
This feature will discourage the object boundary aligning
with contours with small texture change (such as contours
within an object or contours due to illumination).
Edges: We extract edges [11] from the image to measure
their alignment with the edgemap of estimated pose. We
used a modified Chamfer distance from Satkin, et. al. [18].
f
edge
=
1
|R|
X
a2R
min(min
b2I
ka bk,⌧),
1
|I|
X
b2I
min(min
a2R
kb ak,⌧)
(8)
where we use ⌧ 2{10, 25, 50, 1} to control the influence
of outlier edges.
Number of correspondences: f
corr
is a binary vector,
where the i’th dimension indicates if there are more than i
good correspondences between the 3D model and the 2D
image under pose P . Good correspondences are the ones
with local correspondence error (in Eq 2) below a threshold.
2.5. Optimization & Learning
Algorithm 1 Pose set {P } search
For each initial seed pose,
while less than n different candidates found do
Choose a random 3D interest point and its 2D corre-
spondence (this determines a displacement)
for i = 1 to 5 do
Choose a random local correspondence agreeing
with correspondences selected (over i 1 iterations)
end for
Estimate parameters by solving least squares
Find all correspondences agreeing with this solution
Estimate parameters using all correspondences
end while
Our cost function S(P, c) from Eq 1 with G(P ) is a non-
convex function and is not easy to solve directly. Hence,
we first simplify the optimization by quantizing the space
of solutions. Because our L(P, c) is 1 if any local corre-
spondence score is below the threshold, we first find all sets
of correspondences for which all local correspondences are
above the threshold. Then, we find the pose P that mini-
mizes L(P, c)+w
D
(P, c). Finally, we optimize the cost
function in the discrete space by evaluating all candidates.
We use RANSAC in populating a set of candidates by
optimizing L(P, c). Our RANSAC procedure is shown in
Alg 1. We then minimize L( P, c)+w
D
D(P, c) by esti-
mating pose P for each found correspondence c. Given
a set of correspondences c, we estimate pose P using the
Levenberg-Marquardt algorithm minimizing D(P, c).
We again leverage the discretized space from Alg 1 in
order to learn weights for Eq 1. For the subset of {(P, c)}
in the training set, we extract the geometric distance and
global alignment features,
⇥
D(P, c),G(P )
⇤
, and the binary
labels based on the distance from the ground truth (defined
in Eq 9) . Then, we learn the weights using a linear SVM
classifier.
3. Evaluation
3.1. Dataset
Figure 4. Labeling tool: our labeling tool enables users to browse
through 3D models and label point correspondences to an image.
The tool provides a feedback by rendering estimated pose on the
image and an user can edit more correspondences.
In order to develop and evaluate fine pose estimation
based on 3D models, we created a new dataset of images
and 3D models representing typical indoor scenes. We ex-
plicitly collected IKEA 3D models from Google 3D Ware-
house, and images from Flickr. The key difference of this
dataset from previous works [18, 20] is that we align ex-
act 3D models with each image, whereas others provided
coarse pose information without using exact 3D models.
For our dataset, we provide 800 images and 225 3D
models. All 800 images are fully annotated with 90 dif-
(a) 3D models
IKEA object
IKEA room
(b) Aligned Images
Figure 5. Dataset: (a) examples of 3D models we collected from
Google Warehouse, and (b) ground truth images where objects are
aligned with 3D models using our labeling tool. For clarity, we
show only one object per image when there are multiple objects.
ferent models. Also, we separate the data into two differ-
ent splits: IKEAobject and IKEAroom. IKEAobject is
the split where 300 images are queried by individual object
name (e.g. ikea chair poang and ikea sofa ektorp). Hence, it
tends to contain only a few objects at relatively large scales.
IKEAroom is the split where 500 images are queried by
ikea room and ikea home; and contains more complex scene
where multiple objects appear at a smaller scale. Figure 5ab
show examples of our 3D models and annotated images.
For alignment, we created an online tool that allows a
user to browse through models and label point correspon-
dences (usually 5 are sufficient), and check the model’s esti-
mated pose as the user labels. Given these correspondences,
we solve the least square problem of Eq 4 using Levenberg-
Marquardt. Here, we obtain the full intrinsic/extrinsic pa-
rameters except the skewness and principal points. Figure 4
shows a screenshot of our tool.
3.2. Error Metric
We introduce a new error metric for fine-pose estima-
tion. Intuitively, we use the average 3D distance between
all points in the ground truth and the proposal. When the
distance is small, this is close to the average error in view-
ing angle for all points. Formally, given an estimated pose
P
e
and a ground truth pose P
gt
of image I, we obtain corre-
sponding 3D points in the camera space. Then, we compute
the average pair-wise distance between all corresponding
points divided by their distance to the camera. We consider
P is correct if this average value is less than a threshold.
score(P
e
,P
gt
)=
P
X
i
kE
e
X
i
E
gt
X
i
k
2
P
X
i
kE
gt
X
i
k
2
(9)
4. Results
4.1. Correspondences
First of all, we evaluate our algorithm on finding good
correspondences between a 3D model and an image. This
is crucial for the rest of our system as each additional poor
correspondence grows the search space of RANSAC expo-
nentially.
0 50 100 150
0
1
2
3
4
5
6
7
8
9
10
Number of point detections per detector
Detectected keypoints
Ours
Harris
Figure 6. Correspondence evaluation: we are comparing corre-
spondences between our interest point detector and Harris detec-
tor. The minimum number of interest points we need for reliable
pose estimation is 5. Ours can recall 5 correct correspondences
by considering only the top 10 detections per 3D interest point,
whereas the Harris detector requires 100 per point. This results in
effectively 10
5
times fewer search iterations in RANSAC.
