Towards Human Pose Semantic Synthesis in 3D based on Query
Keywords
Mo’taz Al-Hami and Rolf Lakaemper
Department of Computer & Information Sciences, Temple University, Philadelphia, PA 19122, U.S.A.
Keywords:
Human-pose, 2D Human-pose Estimation, 3D Human-pose Reconstruction, Silhouette Value, Hierarchical
Clustering.
Abstract:
The work presented in this paper is part of a project to enable humanoid robots to build a semantic understand-
ing of their environment adopting unsupervised self-learning techniques. Here, we propose an approach to
learn 3-dimensional human-pose conformations, i.e. structural arrangements of a (simplified) human skeleton
model, given only a minimal verbal description of a human posture (e.g. "sitting", "standing", "tree pose").
The only tools given to the robot are knowledge about the skeleton model, as well as a connection to the
labeled images database "google images". Hence the main contribution of this work is to filter relevant results
from an images database, given a human-pose specific query words, and to transform the information in these
(2D) images into a 3D pose that is the most likely to fit the human understanding of the keywords. Steps
to achieve this goal integrate available 2D human-pose estimators using still images, clustering techniques to
extract representative 2D human skeleton poses, and the 3D-pose from 2D-pose estimation. We evaluate the
approach using different query keywords representing different postures.
1 INTRODUCTION
Humanoid robots are increasingly adapting to phys-
ically mimic human actions and poses, which re-
quire an understanding of human poses and how they
can be captured with relation to the robot’s joint
movements. While the translation of a known tar-
get pose to (rotational) joint positioning is a gener-
ally solved problem in inverse kinematics, this paper
aims to generate a target pose description, given a
simple verbal description ("stand", "warrior pose"). It
therefore aims at adding semantic content to a given
query words and their relation to the robot’s skele-
ton model, which represents the robot’s physical self-
understanding. In short, this approach enables the
robot to experience the posture related meaning of
words descriptions. Other than supervised learning
settings, e.g. simulations of children’s training and
learning (Ikemoto et al., 2012), this project is based
on unsupervised self-learning. While the (Ikemoto
et al., 2012) project focuses on improving the interac-
tion with human counterpart, our project is more prag-
matic in the sense that it should quickly and efficiently
enable the robot to understand its environment. With
such a goal in mind, extending the humanoid robot’s
capability to support self-learning is, for obvious rea-
sons (no supervision needed, no ground truth labeled
data needed), an important step.
Recent works (Mu and Yin, 2010), (Jokinen and
Wilcock, 2014) started to improve Internet-based spo-
ken dialogue between a human and a humanoid robot.
The humanoid robot’s side in this conversation uses
an active source, Wikipedia, to improve spoken dia-
logue capabilities. This kind of research opens the
door to think about a different data source, e.g. im-
ages, to act as the basic descriptor for posture relating
tasks. The motivation for using images is their abil-
ity to provide a visual description of a query’s key-
words. This style of analyzing queries could also han-
dle conceptual issues for many applications like (Mu
and Yin, 2010), (Jokinen and Wilcock, 2014).
For non-rigid objects like human-pose, it is im-
portant to understand the relation between the search
query (i.e. keywords representing a human-pose con-
formation) and its related pose conformation, such
that there is a clear description for this query key-
words using images. For example, we want the robot
to learn the posture "standing pose" without any hu-
man intervention. The robot is only allowed to query
a general 2D images database, here google images.
From the image results that google returns from the
query "standing pose", the robot needs to decide, what
420
Al-Hami M. and Lakaemper R..
Towards Human Pose Semantic Synthesis in 3D based on Query Keywords.
DOI: 10.5220/0005258704200427
In Proceedings of the 10th International Conference on Computer Vision Theory and Applications (VISAPP-2015), pages 420-427
ISBN: 978-989-758-091-8
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
the skeleton expression of "standing pose" is. Apply-
ing the search on a "standing pose" query would re-
trieve a group of images which are strongly related to
the query "standing pose", a group of images which
are weakly related to the query, and a group of out-
liers (false positive images) which are completely un-
related. The question now is how to capture the query
related images from the set of all retrieved images.
Next, how to establish a single general pose from
these images which is able to capture the main prop-
erties of the "standing pose" keywords.
In this paper, we propose a framework which
uses query keywords related to specific human-
pose conformation as our input and returns back
a 3D human-pose as output. The importance of this
work is its ability to bridge the gap between estimat-
ing the human pose in 2D images and constructing 3D
human pose from 2D image. To clarify and summa-
rize the approach:
After using a query keywords as an input to
google, and working on the retrieved images, the pro-
posed framework focuses on processing all the re-
trieved images and extract a reliable poses from these
images. Second, based on the extracted poses, ap-
proximate one general pose as a representative pose
for the query. Finally, construct a 3D human pose re-
lated to the approximated 2D pose. This kind of self
awareness allows humanoid-robots to perform a self-
motivated tasks rather than predefined ones.
Our main contributions in this work can be sum-
marized as follows: (1) Propose a hierarchical binary
clustering approach intended to filter poses iteratively
in order to get a consistent and representative clus-
ter of poses. (2) Approximate the representative clus-
ter of poses using one general pose summarizing the
query keywords semantic. The paper is organized
as follows. Related work is described in Section 2.
Human-pose estimation process is discussed in Sec-
tion 3. Extracted human-poses clustering and evalu-
ation are discussed in Section 4 and 5. Finally, con-
clusion remarks appear in Section 6.
2 RELATED WORK
Human-pose consists of a set of related parts, and
these parts are arranged according to a valid hierar-
chical structure. Human-pose estimation in still im-
ages focuses on how to detect a pose and localize its
parts within a 2D image. Many works like (Ramanan
and Sminchisescu, 2006), (Ramanan, 2006), (Fergus
et al., 2003), (Ioffe and Forsyth, 2001), (Ferrari et al.,
2008) use deformable model to create part based tem-
plates, and discover the relationship between parts.
The power of this approach comes from its ability to
detect pose without a prior knowledge about the back-
ground or appearance (i.e. clothes, skin).
Shape matching has been applied in (Gavrila,
2000) to generate a shape template using a distance
transform approach. This template is intended to cap-
ture the objective shape variability. In (Ren et al.,
2005) pairwise constraints between human-pose parts
supported with an image segmentation approach is
used. To locate the candidate parts in a 2D image,
a bottom-up approach is applied.
In (Mori and Malik, 2002), the work provides an
attempt to construct a 3D human-pose model. In addi-
tion to an unseen 2D image, the approach uses a set of
labeled examplers representing different viewpoints
with respect to the used camera. Next, the approach
finds a sufficient match for the unseen image in the
examplers set, and then a 3D model is constructed
based on the labels of these sufficient examplers. The
work in (Yao and Fei-Fei, 2010) focuses on studying
activities that include human object interaction like
many sport activities (i.e. tennis, football). The work
proposes a mutual context approach between human-
pose and objects. This mutual context facilitates the
recognition process of an object and the estimation of
a human-pose as well.
Mixed Bayesian model has been used in (Lan and
Huttenlocher, 2004). The approach consists of hid-
den Markov model and pictorial structure to estimate
human-pose. Since human-pose has a high symme-
try between limbs (i.e. right arm with left arm, and
right leg with left leg), work in (Lan and Huttenlocher,
2005) incorporates this symmetry with tree structure
by taking into account balancing symmetric limbs.
In (Eichner et al., 2012), the work focuses on detect-
ing upper body parts and uses the deformable model
with extension to reduce the search space for parts
localization. The work uses some prior assumption
about the human-pose. These assumptions describe
the relation between the head and the torso and their
upright appearance.
For more human and robot interaction, work
in (Jokinen and Wilcock, 2014) improves social hu-
man robot communication through using a WikiTalk
application. This application provides a social con-
versation system aided with Internet based capability
(i.e. connected to Wikipedia). The humanoid robot
NAO was used as a testing platform for the Wik-
iTalk and improved with gesturing movements. The
work in (Al-Hami and Lakaemper, 2014) applies the
genetic algorithm on NAO humanoid robot pose to
fit well on unknown sittable objects including (boxes
and balls).
TowardsHumanPoseSemanticSynthesisin3DbasedonQueryKeywords
421
Figure 1: Deva Ramanan pictorial structure deformable model approach for human-pose estimation (Ramanan, 2006). (a)
original image, (b) constructed edge map from the original image, (c) Spatial priors (templates) are used based on the learned
spatial priors from a database of labeled joints images, (d) initial parsed model for the whole pose, and (e) initial parsed model
for independent joints.
(a) Standing (b) Sitting (c) Standing (d) warrior
Figure 2: Some examples of human-pose estimation results
when applying approach in (Ferrari et al., 2008) on retrieved
images by google for different pose conformations. The
labels under figures refer to the used query keywords for
these images.
3 HUMAN-POSE ESTIMATION
Human-pose consists of a non-rigid articulated object
having a large number of degrees of freedom. Such
pose structure (shown in Table 1) leads to a high di-
mensional space of possible pose conformations. Pic-
torial structure introduced by Martin Fischler in (Fis-
chler and Elschlager, 1973) has been used to represent
human-pose parts model. The approach takes into ac-
count the spatial conformation and the parts appear-
ance of a pose within an image. For the sake of clar-
ity, assume that the human-pose L consists of m joints
L = {l
1
,l
2
,...,l
m
} where l
i
is the location of i
th
joint in
that pose. Each joint in the model has a specific con-
formation describes its appearance in a 2D image and
identified by the conformation triple [x
i
,y
i
,θ
i
]. (x
i
,y
i
)
represents the location and θ
i
is the orientation of that
joint. The goal of such a model is to detect a human-
pose/poses in an unseen 2D image, then to find each
part in each pose appears in that image.
3.1 Spatial Priors
The process of recognizing a human-pose in a 2D im-
age requires a knowledge about each joint exists in
this pose conformation. Machine learning approaches
have been applied on a set of labeled images including
human-poses. This set of images is labeled human-
pose parts. The task is to learn the localization of
each part in a pose such that the localization should
satisfy some criteria like maximum likelihood (ML)
solution (Ramanan, 2006). The output of this learn-
ing process is a general spatial prior template (local
information template describes part appearance in an
image and its localization) for the human-pose like
the one shown in Fig. 1(c).
In the spatial priors model, a human-pose uses
parameters which capture two main aspects in the
model, spatial relationship S, and appearance A. For
such model M = (S,A), the required parameters are
learned using machine learning approaches using a
data set of labeled images. The process of build-
ing spatial priors (i.e. template for each joint in the
human-pose which favors a certain triple [x
i
,y
i
,θ
i
]) is
concerned with how to determine the P(L) as a prior
distribution.
3.2 Iterative Parsing Approach
Iterative parsing approach proposed by Deva Ra-
manan (Ramanan, 2006) is a general approach (i.e.
does not depend on specific features like clothes,
color, or skin) which is able to recognize a human-
pose’s joints in a 2D image. The contribution in this
approach is its ability to detect and localize a pose’s
joints using iterative parsing process applied on a 2D
image. Throughout each iteration a better features can
VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications
422
be learned and tuned to a specific 2D image, resulting
in an improved pose detection and joints localization.
In the first parse iteration, a spatial priors are used
to produce soft (i.e. roughly) estimate for a pose
joints. This estimate is rough since the spatial priors
are trained on a general labeled database of 2D im-
ages. After the first parsing iteration, the spatial priors
starts to be tuned towards the current image features
(i.e. the iterative parsing approach tune the general
template to more specific one related to the current
image).
The approach in Fig. 1 focuses on building joint-
based region model which determines each human-
pose joint location. In the first iteration, edge map is
used with the spatial priors to generate the initial soft
parse, after that the regions generated in each parse
iteration are used to enhance the pose estimation pro-
cess in the next iteration.
For estimating human-pose joints, the approach
emphasizes the importance of the conditional likeli-
hood model, rather than the classical maximum like-
lihood model. For conditional likelihood the goal is:
max P(L\I, Θ) (1)
which means maximizing the likelihood estimate that
produces the best estimate when applied to labeling a
human-pose parts in a given 2D image I. Based on a
group of labeled training images, the learned model Θ
focuses on learning the model parameters which sat-
isfy this conditional likelihood model. Assuming the
relation between a pose’s joints can be represented as
a tree structure where there is a relation between each
joint and its parent. Using such structure, the log-
linear deformable model can be represented by the
conditional likelihood model:
P(L\I,Θ) exp(
i, jE
ψ(l
i
,l
j
) +
i
φ(l
i
)) (2)
In this model, ψ(l
i
,l
j
) shows the spatial prior for joint
l
i
using its relative arrangement to joint l
j
which both
have a connection in the tree structure. This prior
captures the joint spatial arrangement. The term φ(l
i
)
captures appearance prior for each joint, and serves as
a local image evidence for these joints. For more de-
tails about this log-linear deformable model see (Fer-
rari et al., 2008), (Ramanan and Sminchisescu, 2006).
3.3 Poses Skeletons Extraction
Applying the iterative parsing approach as mentioned
earlier on the retrieved images would produce a
database of extracted poses skeletons. Many im-
plausible skeletons are excluded from the database
Table 1: The used tree structure for representing a human-
pose.
Number Joint Parent
01 Torso -
02 Left Upper Arm Torso
03 Right Upper Arm Torso
04 Left Upper Leg Torso
05 Right Upper Leg Torso
06 Left Lower Arm Left Upper Arm
07 Right Lower Arm Right Upper Arm
08 Left Lower Leg Left Upper Leg
09 Right Lower Leg Right Upper Leg
10 Head Torso
(a) Pose Structure (b) Feature Space
Figure 3: (a): Human-pose structure used in this work. (b):
The used distances which appended in the feature vector.
(we kept only poses where the distances between a
pose joints according to the tree structure is less than
distance σ threshold). Applying the pose estimator
on the retrieved images revealed many limitations.
Some examples from the retrieved results are shown
in Fig. 2. We noticed some wrong segmentation (sep-
arating the human-pose from the background in a 2D
image) for many images (see 2(b)). Another limita-
tion happens in a human-pose estimation even with
correct segmentation, where the estimator localizes
wrongly some of a pose joints (see 2(c)). However,
the estimator was able to segment and estimate many
poses correctly like ones shown in 2(a), 2(d). Since
human-poses in the retrieved images appear at differ-
ent depths and different scales, all extracted human-
poses are scaled to a one unified bounding box.
4 EXTRACTED HUMAN-POSES
CLUSTERING
The poses extraction process suffers from limitations
mentioned earlier. To address such problems, a hier-
archical binary clustering approach is used to extract
a sufficient representative set of human-poses skele-
TowardsHumanPoseSemanticSynthesisin3DbasedonQueryKeywords
423
(a) warrior Pose PCA (b) Standing Pose PCA (c) Tree Pose PCA (d) Sitting Pose PCA
(e) WarriorII Pose Variations. (f) Standing Pose variations. (g) Tree Pose Variations. (h) Sitting Pose Variations.
Figure 4: Poses appearance in space after scaling to a predefined bounding box. (a)-(d) show the plot of the first three principal
components of the poses after transforming the poses space into a new space using PCA analysis. (e)-(h) show a common
typical variation for each pose in the real life.
tons. The basic idea depends on the existing consis-
tency between poses skeletons which are strongly re-
lated to the query keywords. Using a consistency as a
criteria for the selected poses skeletons, we can focus
our attention towards finding a consistent representa-
tive set from human-poses skeletons. Since the avail-
able database (i.e. retrieved human-pose skeletons)
are high dimensional, the consistency is coupled with
poses spatial localization in the space. This spatial
localization in the space depends on the used feature
vector structure (described later).
In Fig. 4 we used principal component analysis
(PCA) to show the database of human-poses in the
space for different pose conformations as well as their
typical variation in real life. By plotting the first three
principal components of the used feature vector, the
localization of poses shows some areas where there
are a high density spots of poses while others are
spread out randomly.
Highlighting the cluster consistency between the
cluster members can offer clues about the main char-
acteristics of a pose conformation. By using a suit-
able feature vector (described later) and using the
iterative hierarchical binary clustering approach us-
ing kmeans clustering agorithm, we can cluster the
poses skeletons. The clustering approach separates
the poses points iteratively into two clusters, one of
them (which has a higher consistency value) is se-
lected to be the cluster of interest while the other one
is discarded. The clustering procedure is repeated
multiple times on the selected cluster until the clus-
ter consistency value (Silhouette value) is more than
a threshold t (see Fig. 5). Obtaining a consistent rep-
resentative cluster provides a subset of related real
poses.
Figure 5: Hierarchical binary clustering approach. At each
level, poses skeletons are divided into two clusters, one of
them which is more consistent is selected while the other
one is discarded. The process is repeated until the repre-
sentative cluster consistency exceeds a threshold value t.
Kmeans clustering algorithm is used in this approach.
4.1 Feature Vector
Kmeans clustering approach separates data points
(poses) based on their proximity to clusters centroids.
The used feature vector structure for representing a
pose is described in Table 2. The feature vector con-
sists of two main parts. The first part keeps center
points of a pose joints as shown in Fig. 3(a) while
the second part keeps distances between a pose joints
as shown in Fig. 3(b). The motivation of using such
structure can be described as follows. The first part
captures a pose joints absolute spatial location using
center points, while the second part captures a pose
joints locations relative to their parent joints and their
peer joints as well.
VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications
424
Table 2: Structure of the used feature vector consists of two parts, one of them captures the spatial localization (center points)
while the other one captures the joints relations (distances). Numbers in the first part refer to center points of specified joints
(i.e. x offset and y offset of a center point). d[a,b] in the second part refers to the Euclidian distance between center points of
joints a and b.
Feature vector structure
Absolute spatial arrangement Relative spatial arrangement
Center points Distances
01 02 03 04 05 06 07 08 09 10 d[2,3] d[6,7] d[2,6] d[3,7] d[4,5] d[8,9] d[4,8] d[5,9]
4.2 Silhouette Value
Silhouette value analysis measures a cluster quality
and consistency. In this analysis each pose in a cluster
is assigned a normalized similarity value (S
i
). This
value reflects a pose similarity to poses in the same
cluster when compared to poses in the other cluster.
S
i
= (b
i
a
i
)/max(a
i
,b
i
) (3)
The similarity is captured through average Euclid-
ian distance from a given pose to all other poses in the
same cluster (a
i
). The other criteria which is the aver-
age Euclidian distance from a pose to all poses from
the other cluster (b
i
). Silhouette values range is be-
tween [-1,1]. To assign a cluster a single silhouette
value, we use the average of all poses silhouette val-
ues in that cluster as representative silhouette value.
The resulted value is used as the criteria in the hier-
archical binary clustering. It decides which cluster is
the relevant one and which cluster has to be discarded.
4.3 Approximate Pose Skeleton Model
Model approximation captures the general character-
istics of a cluster. By approximation, each joint in
the approximate pose skeleton represents the median
orientation of that joint of all poses within a cluster
independently of other joints. Median is chosen since
average is sensitive to outliers. The approximation
process enables us to track the cluster changes at each
level in the hierarchical binary clustering.
4.4 3D Human-pose Reconstruction
The work in (Ramakrishna et al., 2012) constructs a
3D human-pose using a 2D image with landmarks lo-
cated at specific positions and specifying positions of
a human pose in that image. The approach uses a pro-
jected matching pursuit algorithm in order to mini-
mize the error amount between the original landmarks
in the 2D image and the projected ones from the con-
structed 3D model. Assuming a fixed intrinsic cam-
era parameters, the approach is able to approximate
the extrinsic camera parameters and a 3D model us-
ing a corpus (database) of 3D poses. In this work,
(a) L 01 (b) L 02 (c) L 03 (d) L 04 (e) L 05
Figure 6: Warrior approximation pose at different hierarchi-
cal binary clustering levels.
we use approximate poses generated from representa-
tive clusters to be used for 3D reconstruction. Since
the joints positions (x offset and y offset) are already
known, we can automatically mark the required land-
marks.
5 EVALUATION
In experiments, we try to show the effect of the used
hierarchical binary clustering approach in producing a
representative subset of poses. The approach has been
applied to three different queries. Since the extracted
human-poses from the retrieved images (varies be-
tween 700 to 900 images in our experiments) might
include implausible poses, we filtered the extracted
human-poses to plausible ones (the Euclidean dis-
tance between related joints in the hierarchy is below
that a threshold value). Using the filtered set of poses
the hierarchical binary clustering approach is applied
iteratively. The goal is to cluster poses until we get
a consistent cluster (average cluster silhouette value
is at least 0.8). The first query was "warrior pose".
After applying the hierarchical binary clustering ap-
proach 5 levels, we were able to get a consistent clus-
ter. Fig. 6 shows the approximate pose at each level in
the clustering approach. As far as we perform more
iterations in the clustering approach, the pose confor-
mation moves towards a sufficient representative ap-
proximate pose for the original query keywords "war-
rior pose".
The second query in the experiment was "tree
pose". Following the same approach, after 9 levels of
clustering we were able to reach a consistent cluster.
Fig. 7 shows the approximate pose at each level in the
clustering approach. The approximate pose is able to
TowardsHumanPoseSemanticSynthesisin3DbasedonQueryKeywords
425
(a) Tree pose hierarchical binary clustering Silhouette values (b) warrior pose hierarchical binary clustering Silhouette values
(c) Standing pose hierarchical binary clustering Silhouette values
Figure 10: Average Silhouette value at each level in the hierarchical binary clustering approach for the selected and discarded
clusters for three different queries. The stop condition for the clustering process is an average Silhouette value exceeds 0.8.
(a) L 01 (b) L 03 (c) L 05 (d) L 07 (e) L 09
Figure 7: Tree pose approximation pose at different hierar-
chical binary clustering levels.
(a) L 01 (b) L 03 (c) L 04 (d) L 07 (e) L 09
Figure 8: Standing pose approximation pose at different hi-
erarchical binary clustering levels.
(a) Warrior pose (b) Tree pose (c) Standing pose
Figure 9: Reconstructing 3D human-pose models based on
approximate 2D human-pose models.
capture the common characteristic of a real tree pose.
The third query was "standing pose". After 9 levels
of clustering we got a consistent cluster as shown in
Fig. 8.
In Fig. 10, average silhouette values for the se-
lected and discarded clusters are measured at each
clustering level for queries used in each experiment.
As the silhouette value decreases, the amount of in-
consistent poses appearing within a cluster increases.
The last stage uses obtained results from the hi-
erarchical binary clustering approach and translates
them into an approximate 3D human-pose model.
The motivation is to build a 3D model which allows
us to get an automatic visual translation of a query
keywords of a human-pose conformation. Using the
approach in (Ramakrishna et al., 2012), we applied
the projected matching pursuit algorithm on approx-
imate poses. Fig. 9 shows the obtained 3D human-
pose model for each approximate pose.
A case where this approach does not work well,
is when the iterative parsing approach mentioned in
section 3 produces a weak and inaccurate estimation.
In the case of sitting pose, the joints might have
many overlaps, and this makes the estimation process
weak in a majority of cases. A clear example of the
estimated poses are shown in 2(b). In addition to that,
the variations that might occur in the sitting pose is
also high (person might sit on the floor, on different
types of chairs, or might sit with cross legged). Ap-
plying the hierarchical binary clustering approach on
VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications
426
the estimated poses in this case would not pro-
duce a sufficient representative cluster.
6 CONCLUSION
In this paper, we presented a hierarchical binary clus-
tering approach which bridges the gap between 2D
human-pose estimation in an image, and 3D human-
pose reconstruction from 2D human-pose skeleton.
Using this approach we are able to translate query
keywords representing a human-pose conformation
into an approximate 3D human-pose skeleton. The
work in (Al-Hami and Lakaemper, 2014) uses the ge-
netic algorithm to adjust a humanoid robot pose, so
the robot would fit well on an unknown sittable object.
Extending such approach could be accomplished by
allowing the humanoid robot to adopt self-leaning us-
ing simple verbals describing a specific human-pose.
This style of query analysis allows us to extend the
humanoid robots ability toward self-motivated learn-
ing, allowing them to move forward in many applica-
tions. We discussed a hierarchical binary clustering
approach to extract a consistent representative subset
of human-poses. Silhouette value measurement was
used to capture a cluster consistency, and cost trans-
formation was used to rank poses within a cluster ac-
cording to their closeness to the approximate model.
For future work, we want to improve pose estimation
accuracy in 2D images. Also we want to improve the
3D reconstruction performance by forcing joint valid
rotation ranges, such that the constructed 3D pose is
within a valid joints rotations.
REFERENCES
Al-Hami, M. and Lakaemper, R. (2014). Sitting pose gener-
ation using genetic algorithm for nao humanoid robot.
In IEEE Workshop on Advanced Robotics and its So-
cial Impacts (ARSO), 2014, pages 137–142. IEEE.
Eichner, M., Marin-Jimenez, M., Zisserman, A., and Fer-
rari, V. (2012). 2d articulated human pose estimation
and retrieval in (almost) unconstrained still images.
International Journal of Computer Vision, 99(2):190–
214.
Fergus, R., Perona, P., and Zisserman, A. (2003). Ob-
ject class recognition by unsupervised scale-invariant
learning. In IEEE Computer Society Conference on
Computer Vision and Pattern Recognition (CVPR).
2003, volume 2, pages II–264. IEEE.
Ferrari, V., Marin-Jimenez, M., and Zisserman, A. (2008).
Progressive search space reduction for human pose es-
timation. In IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), 2008, pages 1–8.
IEEE.
Fischler, M. A. and Elschlager, R. A. (1973). The repre-
sentation and matching of pictorial structures. IEEE
Transactions on Computers, 22(1):67–92.
Gavrila, D. (2000). Pedestrian detection from a moving ve-
hicle. In Computer Vision ECCV 2000, pages 37–49.
Springer.
Ikemoto, S., Amor, H. B., Minato, T., Jung, B., and Ishig-
uro, H. (2012). Physical human-robot interaction:
Mutual learning and adaptation. Robotics & Automa-
tion Magazine, IEEE, 19(4):24–35.
Ioffe, S. and Forsyth, D. A. (2001). Probabilistic methods
for finding people. International Journal of Computer
Vision, 43(1):45–68.
Jokinen, K. and Wilcock, G. (2014). Multimodal open-
domain conversations with the nao robot. In Natural
Interaction with Robots, Knowbots and Smartphones,
pages 213–224. Springer.
Lan, X. and Huttenlocher, D. P. (2004). A unified spatio-
temporal articulated model for tracking. In IEEE
Computer Society Conference on Computer Vision
and Pattern Recognition (CVPR), 2004, volume 1,
pages I–722. IEEE.
Lan, X. and Huttenlocher, D. P. (2005). Beyond trees:
Common-factor models for 2d human pose recovery.
In Tenth IEEE International Conference on Computer
Vision (ICCV), 2005, volume 1, pages 470–477. IEEE.
Mori, G. and Malik, J. (2002). Estimating human body con-
figurations using shape context matching. In Com-
puter Vision ECCV 2002, pages 666–680. Springer.
Mu, Y. and Yin, Y. (2010). Human-humanoid robot inter-
action system based on spoken dialogue and vision.
In 3rd IEEE International Conference on Computer
Science and Information Technology (ICCSIT), 2010,
volume 6, pages 328–332. IEEE.
Ramakrishna, V., Kanade, T., and Sheikh, Y. (2012). Re-
constructing 3d human pose from 2d image land-
marks. In Computer Vision ECCV 2012, pages 573–
586. Springer.
Ramanan, D. (2006). Learning to parse images of articu-
lated bodies. In Advances in Neural Information Pro-
cessing Systems, pages 1129–1136.
Ramanan, D. and Sminchisescu, C. (2006). Training de-
formable models for localization. In IEEE Computer
Society Conference on Computer Vision and Pattern
Recognition, 2006, volume 1, pages 206–213. IEEE.
Ren, X., Berg, A. C., and Malik, J. (2005). Recovering
human body configurations using pairwise constraints
between parts. In Tenth IEEE International Confer-
ence on Computer Vision (ICCV), 2005, volume 1,
pages 824–831. IEEE.
Yao, B. and Fei-Fei, L. (2010). Modeling mutual context
of object and human pose in human-object interaction
activities. In IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), 2010, pages 17–24.
IEEE.
TowardsHumanPoseSemanticSynthesisin3DbasedonQueryKeywords
427