Human pose estimation

Leonid Sigal, Disney Research, Pittsburgh

Synonyms

– Articulated pose estimation

– Body configuration recovery

Related Concepts

– Human pose tracking

– People tracking
– Articulated pose tracking
– Body parsing
– People parsing

Definition
Human pose estimation is the process of estimating the configuration of the
body (pose) from a single, typically monocular, image.
Background
Human pose estimation is one of the key problems in computer vision that
has been studied for well over 15 years. The reason for its importance is the
abundance of applications that can benefit from such a technology. For example,
human pose estimation allows for higher level reasoning in the context of human-
computer interaction and activity recognition; it is also one of the basic building
blocks for marker-less motion capture (MoCap) technology. MoCap technology
is useful for applications ranging from character animation to clinical analysis
of gait pathologies.
Despite many years of research, however, pose estimation remains a very
difficult and still largely unsolved problem. Among the most significant chal-
lenges are: (1) variability of human visual appearance in images, (2) variability
in lighting conditions, (3) variability in human physique, (4) partial occlusions
due to self articulation and layering of objects in the scene, (5) complexity of
human skeletal structure, (6) high dimensionality of the pose, and (7) the loss
of 3d information that results from observing the pose from 2d planar image
projections. To date, there is no approach that can produce satisfactory results
in general, unconstrained settings while dealing with all of the aforementioned
challenges.
Theory and Application
Human pose estimation is typically formulated probabilistically to account
for ambiguities that may exist in the inference (though there are notable ex-
ceptions, e.g. [11]). In such cases, one is interested in estimating the posterior
Fig. 1. Skeleton Representation: Illustration of the 3d and 2d kinematic tree
skeleton representation on the left and right, respectively.

distribution, p(x|z), where x is the pose of the body and and z is a feature set
derived from the image. The key modeling choices that affect the inference are:

– The representation of the pose – x

– The nature and encoding of image features – z
– The inference framework required to estimate the posterior – p(x|z)

Next, the primary lines of research in pose estimation with respect to these
modeling choices are reviewed. It is worth noting that these three modeling
choices are not always independent. For example, some inference frameworks
are specifically designed to utilize a given representation of the pose.

Representation: The configuration of the human body can be represented in

a variety of ways. The most direct and common representation is obtained by pa-
rameterizing the body as a kinematic tree (see Figure 1), x = {τ, θτ , θ1 , θ2 , ..., θN },
where the pose is encoded using position of the root segment (to keep the kine-
matic tree as short as possible, the pelvis is typically used as the root segment),
τ , orientation of the root segment in the world, θτ , and a set of relative joint
angles, {θi }Ni=1 , that represent the orientations of body parts with respect to
their parents along the tree (e.g., the orientation of the thigh with respect to
the pelvis, shin with respect to the thigh, etc.).
Kinematic tree representation can be obtained for 2d, 2.5d, and 3d body
models. In 3d, τ ∈ R3 and θτ ∈ SO(3); θi ∈ SO(3) for spherical joints (e.g.,
neck), θi ∈ R2 for saddle joints (e.g., wrist), and θi ∈ R1 for hinge joints (e.g.,
knee) and represents the pose of the body in the world. Note, the actual repre-
sentation of the rotations in 3d is beyond the scope of this entry. In 2d, τ ∈ R2
and θτ ∈ R1 ; θi ∈ R1 corresponds to pose of the cardboard person in the image
plane. 2.5d representations are the least common and are extensions of the 2d
representation such that the pose, x, is augmented with (typically discrete) vari-
ables encoding the relative depth (layering) of body parts with respect to one
Fig. 2. Image Features: Illustration of common image features and encoding
methods used in the literature.

another in the 2d cardboard model. In all cases, be it in 2d or 3d, this represen-

tation results in a high-dimensional pose vector, x, in R30 – R70 , depending on
the fidelity and exact parameterization of the skeleton and joints. Alternatively,
one can parameterize the pose of the body by 2d or 3d locations of the major
joints [6]. For example, x = {p1 , p2 , ..., pN }, where pi is the joint location in
the world, pi ∈ R3 , or in an image, pi ∈ R2 . This latter representation is less
common, however, because it is not invariant to the morphology (body segment
lengths) of a given individual.
A typical alternative to the kinematic tree models is to model the body as
a set of parts, x = {x1 , x2 , ..., xM }, each with its own position and orientation
in space, xi = {τi , θi }, that are connected by a set of statistical or physical con-
straints that enforce skeletal (and sometimes image) consistency. Because part-
based parameterization is redundant, it results in an even higher-dimensional
representation. However, it does so in a way that makes it efficient to infer the
pose, as will be discussed in a later section. Methods that utilize such a parame-
terization are often called part-based. As in kinematic tree models, the parts can
be defined in 2d [2,4,7,8,11,16] or in 3d [21], with 2d parameterizations being
significantly more common. In 2d, each part’s representation is often augmented
with an additional variable, si , that accounts for uniform scaling of the body
part in the image, i.e., xi = {τi , θi , si } with τi ∈ R2 , θi ∈ R1 and si ∈ R1 .

Image features: Performance of any pose estimation approach depends sub-

stantially on the observations, or image features, that are chosen to represent
salient parts of the image with respect to the human pose. A related and equally
important issue is one of how these features are encoded. In addition to using
different encodings, some approaches propose to reduce the dimensionality of
the resulting feature vectors through vector quantization or bag-of-words repre-
sentations. These coarser representations simplify feature matching, but at the
expense of losing spatial structure in the image. Common features and encoding
methods are illustrated in Figure 2.
Over the years, many features have been proposed by various authors. The
most common features include: image silhouettes [1], for effectively separating
the person from background in static scenes; color [16], for modeling un-occluded
skin or clothing; edges [16], for modeling external and internal contours of the
body; and gradients [5], for modeling the texture over the body parts. Less
common features include shading and focus [14]. To reduce dimensionality and
increase robustness to noise, these raw features are often encapsulated in image
descriptors, such as shape context [1,2,6], SIFT [6] and histogram of oriented gra-
dients [5]. Alternatively, hierarchical multi-level image encodings can be used,
such as HMAX [12], spatial pyramids [12], and vocabulary trees [12]. The ef-
fectiveness of different feature types on pose estimation has been studied in the
context of several inference architectures; see [2] and [12] for discussions and
quantitative analyses.

Inference (regression models): Characterizing the posterior distribution,

p(x|z), can be done in a number of ways. Perhaps the most intuitive way is to
define a parametric [1,6,12] or non-parametric [15,18,22,25] form for the con-
ditional distribution p(x|z) and learn the parameters of that distribution from
a set of training exemplars. This class of models is more widely known as dis-
criminative methods, and they have been shown to be very effective for pose
estimation. Such methods directly learn p(x|z) from a labelled dataset of poses
and corresponding images, D = {(xi , zi )}N
i=1 , which can be produced artificially
using computer graphics software packages (e.g., Poser) [1,12,18]. The inference
takes a form of probabilistic regression. Once a regression function is learned,
a scanning window approach is typically used at test time to detect a portion
of the image (bounding box) where the person resides; p(x|z) is then used to
characterize the configuration of the person in that target window.
The simplest method in this category is the one of linear regression [1], where
the body configuration, x, is assumed to be a linear combination of the image
features, z, with additive Gaussian noise,

x = A[z − µz ] + µx + ν; ν ∼ N (0, Σ);

PN N
µx = N1 i=1 xi and µz = N1 i=1 zi are means computed over the training
P
samples to center the data. Alternatively, this can be written as:

p(x|z) = N (A[z − µz ] + µx , Σ). (1)

The regression coefficients, A, can be learned easily from paired training samples,
D = {(xi , zi )}N
i=1 , using the least squares formulation (see [1] for details).
Parametric vs. Non-parametric: Parametric discriminative methods [1,6,12] are
appealing because the model representation is fixed with respect to the size
of the training dataset D. However, simple parametric models, such as Linear
Regression [1] or Relevance Vector Machine [1], are unable to deal with com-
plex non-linear relationships between image features and poses. Non-parametric
methods, such as Nearest Neighbor Regression [18] or Kernel Regression [18],
are able to model arbitrary complex relationships between input features and
output poses. The disadvantage of these non-parametric methods is that the
model and inference complexity are both functions of the training set size. For
example, in Kernel Regression,

N
X Kz (z, zi )
p(x|z) = Kx (x, xi ) PN , (2)
i=1 k=1 Kz (z, zk )

where Kx (·, ·) and Kz (·, ·) are kernel functions measuring the similarity of the
arguments (e.g., Gaussian kernels), the inference complexity is O(N ) (where N
is the size of the training dataset). More sophisticated non-parametric methods,
such as Gaussian Process Latent Variable Models (GPLVMs), can have even
higher complexity; GPLVMs have O(N 3 ) learning and O(N 2 ) inference com-
plexity. In practice, non-parametric methods tend to perform better but are
slower.
Dealing with ambiguities: If one assumes that p(x|z) is uni-modal [1], conditional
expectation can be used to characterize the plausible configuration of the person
in an image given the learned model. For example, for linear regression in Eq. (1),
E[x|z] = A[z − µz ] + µx ; for kernel regression in Eq. (2),

N
X Kz (z, zi )
E[x|z] = xi PN . (3)
i=1 k=1 Kz (z, zk )

In practice, however, most features under standard imaging conditions are am-
biguous, resulting in multi-modal distributions. Ambiguities naturally arise in
image projections, where multiple poses can result in similar, if not identical, im-
age features (e.g., front- and back-facing poses yield nearly identical silhouette
features). To account for these ambiguities, parametric mixture models were in-
troduced in the form of Mixture of Experts [6,12]. Non-parametric alternatives,
such as Local Gaussian Process Latent Variable Models (LGPLVM) [25], clus-
ter the data into convex local sets and make uni-modal predictions within each
cluster, or search for prominent modes in p(x|z) [15,22].
Learning: Obtaining the large datasets that are required for learning discrim-
inative models that can generalize across motions and imaging conditions is
challenging. Synthetic datasets often do not exhibit the imaging characteristics
present in real images, and real fully-labeled datasets are scarce. Furthermore,
even if large datasets could be obtained, learning from vast amounts of data is
not a trivial task [6]. To address this issue, two solutions were introduced: (1)
learning from small datasets by discovering an intermediate low dimensional la-
tent space for regularization [15,22] and (2) learning in semi-supervised settings,
where a relatively small dataset of paired samples is accompanied by a large
amount of unlabeled data [12,15,22].
Limitations: Despite popularity and lots of successes, discriminative methods
do have limitations. First, they are only capable of recovering a relative 3d
configuration of the body and not its position in 3d space. The reason for this
is practical, as reasoning about position in 3d space would require prohibitively
large training datasets that span the entire 3d volume of the space visible from
the camera. Second, their performance tends to degrade as the distributions of
test and training data start to diverge; in other words, generalization remains
one of the key issues. Lastly, learning discriminative models efficiently from large
datasets that cover wide range of realistic activities and postures remains a
challenging task.
Inference (generative): Alternatively, one can take a generative approach and
express the desired posterior, p(x|z), as a product of a likelihood and a prior:

p(x|z) ∝ p(z|x) p(x) . (4)

| {z } |{z}
likelihood prior

Characterizing this high-dimensional posterior distribution is typically hard;

hence, most approaches rely on a posteriori (MAP) solutions that look for the
most probable configurations that are both typical (have high prior probability)
and can explain the image data well (have high likelihood):

xM AP = arg max p(x|z). (5)

Searching for such configurations, however, in the high-dimensional (40+) ar-

ticulation space is very challenging and most approaches frequently get stuck
in local optima. Global hierarchical search methods, such as Annealed Particle
Filter [10], have shown some promising results for simple skeletal configurations,
where body is mostly upright, and when observations from multiple cameras are
available. For the more general articulations and monocular observations, that
are often the focus of pose estimation algorithms, this class of methods has not
been very successful to date.
Inference (part-based models): To battle the inference complexity of gener-
ative models, part-based models have been introduced. These methods originate
in the object recognition community with formulation of Fischler and Elschlager
(1973) and assume that a body can be represented as an assembly of parts that
are connected by constraints imposed by the joints within the skeletal structure
(and, sometimes, by the image constraints imposed by projections onto an im-
age plane that account for occlusions). This formulation reduces the inference
complexity because likely body part locations can be searched for independently,
only considering the nearby body parts that constrain them, which significantly
prunes the total search space.
Fig. 3. Pictorial Structures Model: Illustrated is the depiction of the 10-part
tree-structured pictorial structures model (middle) and a non-tree-structured
(loopy) pictorial structures model (right). In the non-tree-structured model ad-
ditional constraints encoding occlusions are illustrated in blue.

Among the earliest successes along this line of research is the work of Lee and
Cohen [13]. Their approach focused on obtaining proposal maps for the locations
of individual joints within an image. These proposal maps were obtained based
on a number of features that were computed densely over the image. For example,
face detection was used to obtain hypotheses for the location of the head; head-
shoulder contour matching, obtained using a deformable contour model and
gradient descent, was used as evidence for shoulder joint locations; elliptical
skin regions, obtained using skin-color segmentation, were used to determine
the locations of the lower arms and lower legs. In addition, second-derivative
(ridge) observations were used as evidence for other limbs of the body. Given
proposals for the different joints, weighted by the confidence of corresponding
detectors, a data-driven Markov Chain Monte Carlo (MCMC) approach was
used to recover 3d configurations of the skeleton. This inference relied on direct
inverse kinematics (IK) obtained from 2d proposal maps. To further improve the
results, a kinematic jump proposal process was also introduced. The kinematic
jump proposal process involves flipping a body part or a set of parts (i.e., the
head, a hand, or an entire arm) in the depth direction around its pivotal joint.
Other part-based approaches try to assemble regions of an image into body
parts and successively construct those parts into a body. Prime examples of such
methods are introduced by Mori et al . [14] and Ren et al . [17]. In [14] super-pixels
were first assembled into body parts based on the evaluation of low-level image
cues, including contour, shape, shading, and focus. The part-proposals were then
pruned and assembled together using length, body part adjacency, and clothing
symmetry. A similar approach was taken in [17], but line segments were used
instead of assembling super-pixels. Parallel lines were assembled into candidate
parts using a set of predefined rules, and the candidate parts were in turn as-
sembled into the body with a set of joint, scale, appearance, and orientation
consistency constraints. Unlike [14], the search for the most probable body con-
figurations was formulated as a solution to an Integer Quadratic Programming
(IQP) problem.
 The most traditional and successful approach, however, is to represent the
body using a Markov Random Field (MRF) with body parts corresponding to
the nodes and constraints between parts encoded by potential functions that
account for physical and statistical dependencies (see Figure 3). Formally, the
posterior, p(x|z), can be expressed as:

p(x|z) ∝ p(z|x)p(x)
= p(z|{x1 , x2 , ..., xM })p({x1 , x2 , ..., xM })
M
Y Y
≈ p(z|xi ) p(x1 ) p(xi , xj ) . (6)
i=1 (i,j)∈E
| {z }| {z }
likelihood prior

In this case, pose estimation takes the form of inference in a general MRF net-
work. The inference can be solved efficiently using message-passing algorithms,
such as Belief Propagation (BP). BP consists of two distinct phases: (1) a set of
message-passing iterations are executed to propagate consistent part estimates
within a graph, and (2) marginal posterior distributions are estimated for every
body part [2,8,16]. A typical formulation looks at the configuration of the body
in the 2d image plane and assumes discretization of the pose for each individual
part, e.g., xi = {τi , θi , si } where τi ∈ R2 is the location and θi ∈ R1 and si ∈ R1
are orientation and scale of the part i (represented as a rectangular patch) in the
image plane. As a result, the inference is over a set of discrete part configurations
li ∈ Z (for part i), where Z is the enumeration of poses for a part in an image (li
is a discrete version of xi ). With an additional assumption of pair-wise potentials
that account for kinematic constraints, the model forms a tree-structured graph
known as the Tree-structured Pictorial Structures (PS) model. An approximate
inference with continuous variables is also possible [20,21].
Inference in the tree-structured PS model first proceeds by sending recur-
sively defined messages of the form:
X Y
mi→j (lj ) = p(li , lj )p(z|li ) mk→i (li ), (7)
li k∈A(i)\j

where mi→j is the message from part i to part j, with p(li , lj ) measuring the
compatibility of poses for the two parts and p(z|li ) the likelihood, and A(i) \ j is
the set of parts in the graph adjacent to i except for j. Compatibility, p(li , lj ), is
often measured by the physical consistency of two parts at the joint, or by their
statistical (e.g., angular) co-occurrence with respect to one another. In a tree-
structured PS graph, these messages are sent from the outermost extremities
inward and then back outward.
Once all of the message updates are complete, the marginal posteriors for all
of the parts can be estimated as:
Y
p(li |z) ∝ p(z|li ) mj→i (li ). (8)
j∈A(i)
Similarly, the most likely configuration can be obtained as a MAP estimate:

li,M AP = arg max p(li |z). (9)

li

One of the key benefits of the Pictorial Structures (PS) paradigm is its
simplicity and efficiency. In PS exact inference is possible in the time linear
to the number of discrete configurations a given part can assume. Because of
this property, recent implementations [2] can handle the pixel-dense configura-
tions of parts that result in millions of potential discrete states for each body
part. The linear complexity comes from the observation that a generally com-
plex non-Gaussian prior over neighboring parts, p(xi , xj ), can be expressed as
a Gaussian prior over the transformed locations corresponding to joints, mainly
p(xi , xj ) = N (Tij (xi ); Tji (xj ), Σij ). This is done by defining a transformation,
Tij (xi ), that maps a common joint between parts i and j, defined in the part i’s
coordinate frame, to its location in the image space. Similarly, Tji (xj ) defines the
transformation from the same common joint defined in the j’s coordinate frame
to the location in the image plane. This transformation allows the inference to
use an efficient solution that involves convolution (see [8] for more details).
Performance: Recently, it has been shown that the effectiveness of a PS model
is closely tied to the quality of the part likelihoods [2]. Discriminatively trained
models [16] and more complex appearance models [2] tend to outperform models
defined by hand [8]. Methods that learn likelihood cascades, corresponding to
better and better features tuned to a particular image, have also been explored
for both superior speed and performance [16]. Most recent discriminative formu-
lation of PS model allows joint learning of part appearances and model structure
[26] using structural Support Vector Machine (SVM).
Speed: Cascades of part detectors serve not only to improve performance, but
also to speed up the inference (e.g., [23]). Fast likelihoods can be used to prune
away large parts of the search space before applying more complex and computa-
tionally expensive likelihood models. Other approaches to speed up performance
include data-driven methods (e.g., data-driven Belief Propagation). These meth-
ods look for the parts in an image first and then assemble a small set of the part
candidates into the body (akin to the methods of [14,17]). The problem with
such approaches is that any occluded parts are missed altogether because they
cannot be detected by the initial part detectors. Inference can also be sped up
by using progressive search refinement methods [9]. For example, some methods
use upper body detectors to restrict the search to promising parts of the image
instead of searching the whole image.
Non-tree structured extensions: Although tree-structured PS models are compu-
tationally efficient and exact, they generally are not sufficient to model all the
necessary constraints imposed by the body. More complex relationships among
the parts that fall outside of the realm of these models include non-penetration
constraints and occlusion constraints [20]. Incorporating such relationships into
the model adds loops corresponding to long-range dependencies between body
parts. These loops complicate inference because: (1) no optimal solutions can
be found efficiently (message passing algorithms, like BP, are not guaranteed to
converge in loopy graphs) and (2) even approximate inference is typically com-
putationally expensive. Despite these challenges, it has been argued that adding
such constraints is necessary to improve performance [4]. To alleviate some of the
inference complexities with these non-tree-structured models, a number of com-
peting methods have been introduced. Early attempts used sampling techniques
from the tree-structured posterior as proposals for evaluation of a more complex
non-tree-structured model [7,8]. To obtain optimality guarantees, branch-and-
bound search was recently proposed by Tian et al . [24], with the tree-structured
solutions as a lower bound on the more complex loopy model energy.

Open problems
Despite much progress in the field, pose estimation remains a challenging
and still largely unsolved task. Progress has been made in estimating the con-
figurations of mostly unoccluded and isolated subjects. Open problems include
dealing with multiple, potentially interacting people (e.g., [7]), and tolerance to
unexpected occlusions. Future research is also likely to expand on the types of
postures and imaging conditions that the current algorithms can handle.
To date, most successful pose estimation approaches have been bottom-up.
This observation applies to both discriminative approaches and part-based ap-
proaches. However, it seems short-sighted to assume that the general pose esti-
mation problem can be solved purely in a bottom-up fashion. Top-down infor-
mation may be useful for enforcing global pose consistency, and a combination of
top-down and bottom-up inference is likely to lead to success faster. The recent
success of combining bottom-up part-based models with 3d top-down priors [3]
is encouraging and should be built upon to produce models that can deal with
more complex postures and motions. Earlier attempts at building hierarchical
models [27] may also be worth revisiting with the newfound insights.
Finally, there is significant evidence suggesting that successfully estimating
pose independently at every frame is a very ill-posed problem. Spatio-temporal
models that aggregate information over time [3] are emerging as a way to reg-
ularize performance obtained in individual frames and smooth out the noise in
the estimates. Leveraging all sources of generic prior knowledge, such as spatial
layout of the body and temporal consistency of poses, and rich image observation
models is critical in advancing the state-of-the-art.

Recommended Readings

[1] A. Agarwal and B. Triggs. Recovering 3d human pose from monocular

images. IEEE Transactions on Pattern Analysis and Machine Intelligence,
28(1):44–58, 2006.
[2] M. Andriluka, S. Roth, and B. Schiele. Pictorial structures revisited: People
detection and articulated pose estimation. In IEEE Conference on Com-
puter Vision and Pattern Recognition, 2009.
 [3] M. Andriluka, S. Roth, and B. Schiele. Monocular 3d pose estimation and
tracking by detection. In IEEE Conference on Computer Vision and Pattern
Recognition, 2010.
[4] M. Bergtholdt, J. Kappes, S. Schmidt, and C. Schnorr. A study of parts-
based object class detection using complete graphs. Internation Journal of
Computer Vision, 2010.
[5] L. Bo and C. Sminchisescu. Twin gaussian processes for structured predic-
tion. International Journal of Computer Vision, 2010.
[6] L. Bo, C. Sminchisescu, A. Kanaujia, and D. Metaxas. Fast algorithms for
large scale conditional 3d prediction. In IEEE Conference on Computer
Vision and Pattern Recognition, 2008.
[7] M. Eichner and V. Ferrari. We are family: Joint pose estimation of multiple
persons. In European Conference on Computer Vision, 2010.
[8] P. F. Felzenszwalb and D. P. Huttenlocher. Pictorial structures for object
recognition. International Journal of Computer Vision, 61(1):55–79, 2005.
[9] V. Ferrari, M. J. Marn-Jimnez, and A. Zisserman. Progressive search space
reduction for human pose estimation. In IEEE Conference on Computer
Vision and Pattern Recognition, 2008.
[10] J. Gall, B. Rosenhahn, T. Brox, and H.-P. Seidel. Optimization and filtering
for human motion capture. International Journal of Computer Vision, 87(1–
2):75–92, 2010.
[11] H. Jiang. Human pose estimation using consistent max-covering. In IEEE
International Conference on Computer Vision, 2009.
[12] A. Kanaujia, C. Sminchisescu, and D. Metaxas. Semi-supervised hierar-
chical models for 3d human pose reconstruction. In IEEE Conference on
Computer Vision and Pattern Recognition, 2007.
[13] M. W. Lee and I. Cohen. Proposal maps driven mcmc for estimating human
body pose in static images. In IEEE Conference on Computer Vision and
Pattern Recognition, 2004.
[14] G. Mori, X. Ren, A. Efros, and J. Malik. Recovering human body configu-
rations: Combining segmentation and recognition. In IEEE Conference on
Computer Vision and Pattern Recognition, 2004.
[15] R. Navaratnam, A. Fitzgibbon, and R. Cipolla. The joint manifold model for
semi-supervised multi-valued regression. In IEEE International Conference
on Computer Vision, 2007.
[16] D. Ramanan. Learning to parse images of articulated bodies. In Neural
Information and Processing Systems, 2006.
[17] X. Ren, A. C. Berg, and J. Malik. Recovering human body configurations
using pair-wise constraints between parts. In International Conference on
Computer Vision, 2005.
[18] G. Shakhnarovich, P. Viola, and T. Darrell. Fast pose estimation with pa-
rameter sensitive hashing. In International Conference on Computer Vision,
2003.
[19] L. Sigal, A. Balan, and M. J. Black. Combined discriminative and generative
articulated pose and non-rigid shape estimation. In Neural Information and
Processing Systems, 2007.
[20] L. Sigal and M. J. Black. Measure locally, reason globally:occlusion-sensitive
articulated pose estimation. In IEEE Conference on Computer Vision and
Pattern Recognition, 2006.
[21] L. Sigal, M. Isard, B. H. Sigelman, and M. J. Black. Attractive people:
Assembling loose-limbed models using non-parametric belief propagation.
In Advances in Neural Information Processing Systems, 2003.
[22] L. Sigal, R. Memisevic, and D. J. Fleet. Shared kernel information embed-
ding for discriminative inference. In IEEE Conference on Computer Vision
and Pattern Recognition, 2009.
[23] V. K. Singh, R. Nevatia, and C. Huang. Efficient inference with multi-
ple heterogeneous part detectors for human pose estimation. In European
Conference on Computer Vision, pages 314–327, 2010.
[24] T.-P. Tian and S. Sclaroff. Fast globally optimal 2d human detection with
loopy graph models. In IEEE Conference on Computer Vision and Pattern
Recognition, 2010.
[25] R. Urtasun and T. Darrell. Sparse probabilistic regression for activity-
independent human pose inference. In IEEE Conference on Computer Vi-
sion and Pattern Recognition, 2008.
[26] Y. Yang and D. Ramanan. Articulated pose estimation with flexible
mixture-of-parts. In IEEE Conference on Computer Vision and Pattern
Recognition, 2011.
[27] J. Zhang, J. Luo, R. Collins, and Y. Liu. Body localization in still im-
ages using hierarchical models and hybrid search. In IEEE Conference on
Computer Vision and Pattern Recognition, 2006.