DeepPose: Human Pose Estimation via Deep Neural Networks
Alexander Toshev
[email protected]Google
Christian Szegedy
[email protected]Google
Figure 1. Besides extreme variability in articulations, many of the
joints are barely visible. We can guess the location of the right
arm in the left image only because we see the rest of the pose and
anticipate the motion or activity of the person. Similarly, the left
body half of the person on the right is not visible at all. These
are examples of the need for holistic reasoning. We believe that
DNNs can naturally provide such type of reasoning.
Abstract
We propose a method for human pose estimation based
on Deep Neural Networks (DNNs). The pose estimation
is formulated as a DNN-based regression problem towards
body joints. We present a cascade of such DNN regressors which results in high precision pose estimates. The
approach has the advantage of reasoning about pose in a
holistic fashion and has a simple but yet powerful formulation which capitalizes on recent advances in Deep Learning. We present a detailed empirical analysis with state-ofart or better performance on four academic benchmarks of
diverse real-world images.
1. Introduction
The problem of human pose estimation, defined as the
problem of localization of human joints, has enjoyed substantial attention in the computer vision community. In
Fig. 1, one can see some of the challenges of this problem – strong articulations, small and barely visible joints,
occlusions and the need to capture the context.
The main stream of work in this field has been motivated
mainly by the first challenge, the need to search in the large
space of all possible articulated poses. Part-based models
lend themselves naturally to model articulations ([16, 8])
and in the recent years a variety of models with efficient
inference have been proposed ([6, 18]).
The above efficiency, however, is achieved at the cost of
limited expressiveness – the use of local detectors, which
reason in many cases about a single part, and most importantly by modeling only a small subset of all interactions
between body parts. These limitations, as exemplified in
Fig. 1, have been recognized and methods reasoning about
pose in a holistic manner have been proposed [15, 20] but
with limited success in real-world problems.
In this work we ascribe to this holistic view of human
pose estimation. We capitalize on recent developments of
deep learning and propose a novel algorithm based on a
Deep Neural Network (DNN). DNNs have shown outstanding performance on visual classification tasks [14] and more
recently on object localization [22, 9]. However, the question of applying DNNs for precise localization of articulated
objects has largely remained unanswered. In this paper we
attempt to cast a light on this question and present a simple
and yet powerful formulation of holistic human pose estimation as a DNN.
We formulate the pose estimation as a joint regression
problem and show how to successfully cast it in DNN settings. The location of each body joint is regressed to using
as an input the full image and a 7-layered generic convolutional DNN. There are two advantages of this formulation.
First, the DNN is capable of capturing the full context of
each body joint – each joint regressor uses the full image
as a signal. Second, the approach is substantially simpler
to formulate than methods based on graphical models – no
need to explicitly design feature representations and detectors for parts; no need to explicitly design a model topology
and interactions between joints. Instead, we show that a
generic convolutional DNN can be learned for this problem.
Further, we propose a cascade of DNN-based pose predictors. Such a cascade allows for increased precision of
joint localization. Starting with an initial pose estimation,
based on the full image, we learn DNN-based regressors
which refines the joint predictions by using higher resolution sub-images.
We show state-of-art results or better than state-of-art on
1
four widely used benchmarks against all reported results.
We show that our approach performs well on images of people which exhibit strong variation in appearance as well as
articulations. Finally, we show generalization performance
by cross-dataset evaluation.
2. Related Work
The idea of representing articulated objects in general,
and human pose in particular, as a graph of parts has been
advocated from the early days of computer vision [16]. The
so called Pictorial Strictures (PSs), introduced by Fishler
and Elschlager [8], were made tractable and practical by
Felzenszwalb and Huttenlocher [6] using the distance transform trick. As a result, a wide variety of PS-based models
with practical significance were subsequently developed.
The above tractability, however, comes with the limitation of having a tree-based pose models with simple binary
potential not depending on image data. As a result, research
has focused on enriching the representational power of the
models while maintaining tractability. Earlier attempts to
achieve this were based on richer part detectors [18, 1, 4].
More recently, a wide variety of models expressing complex
joint relationships were proposed. Yang and Ramanan [26]
use a mixture model of parts. Mixture models on the full
model scale, by having mixture of PSs, have been studied
by Johnson and Everingham [13]. Richer higher-order spatial relationships were captured in a hierarchical model by
Tian et al. [24]. A different approach to capture higherorder relationship is through image-dependent PS models,
which can be estimated via a global classifier [25, 19, 17].
Approaches which ascribe to our philosophy of reasoning about pose in a holistic manner have shown limited
practicality. Mori and Malik [15] try to find for each test
image the closest exemplar from a set of labeled images
and transfer the joint locations. A similar nearest neighbor
setup is employed by Shakhnarovich et al. [20], who however use locality sensitive hashing. More recently, Gkioxari
et al. [10] propose a semi-global classifier for part configuration. This formulation has shown very good results on
real-world data, however, it is based on linear classifiers
with less expressive representation than ours and is tested
on arms only. Finally, the idea of pose regression has been
employed by Ionescu et al. [11], however they reason about
3D pose.
The closest work to ours uses convolution NNs together
with Neighborhood Component Analysis to regress toward
a point in an embedding representing pose [23]. However,
this work does not employ a cascade of networks. Cascades
of DNN regressors have been used for localization, however
of facial points [21].
3. Deep Learning Model for Pose Estimation
We use the following notation. To express a pose, we encode the locations of all k body joints in pose vector defined
as y = (. . . , y
T
i
, . . .)
T
, i ∈ {1, . . . , k}, where yi contains
the x and y coordinates of the i
th joint. A labeled image is
denoted by (x, y) where x stands for the image data and y
is the ground truth pose vector.
Further, since the joint coordinates are in absolute image
coordinates, it proves beneficial to normalize them w. r. t. a
box b bounding the human body or parts of it. In a trivial
case, the box can denote the full image. Such a box is defined by its center bc ∈ R
2
as well as width bw and height
bh: b = (bc, bw, bh). Then the joint yi can be translated by
the box center and scaled by the box size which we refer to
as normalization by b:
N(yi
; b) =
1/bw 0
0 1/bh
(yi − bc) (1)
Further, we can apply the same normalization to the elements of pose vector N(y; b) = (. . . , N(yi
; b)
T
, . . .)
T
resulting in a normalized pose vector. Finally, with a slight
abuse of notation, we use N(x; b) to denote a crop of the
image x by the bounding box b, which de facto normalizes
the image by the box. For brevity we denote by N(·) normalization with b being the full image box.
3.1. Pose Estimation as DNN-based Regression
In this work, we treat the problem of pose estimation as
regression, where the we train and use a function ψ(x; θ) ∈
R
2k which for an image x regresses to a normalized pose
vector, where θ denotes the parameters of the model. Thus,
using the normalization transformation from Eq. (1) the
pose prediction y
∗
in absolute image coordinates reads
y
∗ = N
−1
(ψ(N(x); θ)) (2)
Despite its simple formulation, the power and complexity of the method is in ψ, which is based on a convolutional
Deep Neural Network (DNN). Such a convolutional network consists of several layers – each being a linear transformation followed by a non-linear one. The first layer takes
as input an image of predefined size and has a size equal to
the number of pixels times three color channels. The last
layer outputs the target values of the regression, in our case
2k joint coordinates.
We base the architecture of the ψ on the work by
Krizhevsky et al. [14] for image classification since it has
shown outstanding results on object localization as well
[22]. In a nutshell, the network consists of 7 layers (see
Fig. 2 left). Denote by C a convolutional layer, by LRN
a local response normalization layer, P a pooling layer
and by F a fully connected layer. Only C and F layers
(xi, yi) (x(s-1)i, y (s-1) i)
xs
i - x(s-1)i
ys
i - y(s-1)i
Initial stage Stage s
send refined values
to next stage
220 x 220
DNN-based regressor
27 x 27 x 128
13 x 13 x 192
13 x 13 x192
13 x 13 x192
4096
4096
55 x 55 x 48
xi
yi ...
DNN-based refiner
27 x 27 x 128
13 x 13 x 192
13 x 13 x192
13 x 13 x192
4096
4096
55 x 55 x 48
Figure 2. Left: schematic view of the DNN-based pose regression. We visualize the network layers with their corresponding dimensions,
where convolutional layers are in blue, while fully connected ones are in green. We do not show the parameter free layers. Right: at stage
s, a refining regressor is applied on a sub image to refine a prediction from the previous stage.
contain learnable parameters, while the rest are parameter free. Both C and F layers consist of a linear transformation followed by a nonlinear one, which in our case
is a rectified linear unit. For C layers, the size is defined as width × height × depth, where the first two dimensions have a spatial meaning while the depth defines
the number of filters. If we write the size of each layer in
parentheses, then the network can be described concisely
as C(55 × 55 × 96) − LRN − P − C(27 × 27 × 256) −
LRN − P − C(13 × 13 × 384) − C(13 × 13 × 384) −
C(13 × 13 × 256) − P − F(4096) − F(4096). The filter
size for the first two C layers is 11 × 11 and 5 × 5 and for
the remaining three is 3 × 3. Pooling is applied after three
layers and contributes to increased performance despite the
reduction of resolution. The input to the net is an image
of 220 × 220 which via stride of 4 is fed into the network.
The total number of parameters in the above model is about
40M. For further details, we refer the reader to [14].
The use of a generic DNN architecture is motivated by
its outstanding results on both classification and localization
problems. In the experimental section we show that such a
generic architecture can be used to learn a model resulting
in state-of-art or better performance on pose estimation as
well. Further, such a model is a truly holistic one — the
final joint location estimate is based on a complex nonlinear
transformation of the full image.
Additionally, the use of a DNN obviates the need to design a domain specific pose model. Instead such a model
and the features are learned from the data. Although the regression loss does not model explicit interactions between
joints, such are implicitly captured by all of the 7 hidden
layers – all the internal features are shared by all joint regressors.
Training The difference to [14] is the loss. Instead of a
classification loss, we train a linear regression on top of the
last network layer to predict a pose vector by minimizing
L2 distance between the prediction and the true pose vector. Since the ground truth pose vector is defined in absolute image coordinates and poses vary in size from image to
image, we normalize our training set D using the normalization from Eq. (1):
DN = {(N(x), N(y))|(x, y) ∈ D} (3)
Then the L2 loss for obtaining optimal network parameters
reads:
arg min
θ
X
(x,y)∈DN
X
k
i=1
||yi − ψi(x; θ)||2
2
(4)
For clarity we write out the optimization over individual
joints. It should be noted, that the above objective can
be used even if for some images not all joints are labeled.
In this case, the corresponding terms in the sum would be
omitted.
The above parameters θ are optimized for using Backpropagation in a distributed online implementation. For
each mini-batch of size 128, adaptive gradient updates are
computed [3]. The learning rate, as the most important parameter, is set to 0.0005. Since the model has large number
of parameters and the used datasets are of relatively small
size, we augment the data using large number of randomly
translated image crops (see Sec. 3.2), left/right flips as well
as DropOut regularization for the F layers set to 0.6.
3.2. Cascade of Pose Regressors
The pose formulation from the previous section has the
advantage that the joint estimation is based on the full image and thus relies on context. However, due to its fixed
input size of 220 × 220, the network has limited capacity
to look at detail – it learns filters capturing pose properties
at coarse scale. These are necessary to estimate rough pose
but insufficient to always precisely localize the body joints.
Note that we cannot easily increase the input size since this
will increase the already large number of parameters. In order to achieve better precision, we propose to train a cascade
of pose regressors. At the first stage, the cascade starts off
by estimating an initial pose as outlined in the previous section. At subsequent stages, additional DNN regressors are
trained to predict a displacement of the joint locations from
previous stage to the true location. Thus, each subsequent
stage can be thought of as a refinement of the currently predicted pose, as shown in Fig. 2.
Further, each subsequent stage uses the predicted joint
locations to focus on the relevant parts of the image – subimages are cropped around the predicted joint location from
previous stage and the pose displacement regressor for this
joint is applied on this sub-image. In this way, subsequent
pose regressors see higher resolution images and thus learn
features for finer scales which ultimately leads to higher
precision.
We use the same network architecture for all stages of
the cascade but learn different network parameters. For
stage s ∈ {1, . . . , S} of total S cascade stages, we denote by θs the learned network parameters. Thus, the
pose displacement regressor reads ψ(x; θs). To refine a
given joint location yi we will consider a joint bounding
box bi capturing the sub-image around yi
: bi(y; σ) =
(yi
, σdiam(y), σdiam(y)) having as center the i-th joint
and as dimension the pose diameter scaled by σ. The diameter diam(y) of the pose is defined as the distance between
opposing joints on the human torso, such as left shoulder
and right hip, and depends on the concrete pose definition
and dataset.
Using the above notation, at the stage s = 1 we start with
a bounding box b
0 which either encloses the full image or
is obtained by a person detector. We obtain an initial pose:
Stage 1 : y
1 ← N
−1
(ψ(N(x; b
0
); θ1); b
0
) (5)
At each subsequent stage s ≥ 2, for all joints i ∈ {1, . . . , k}
we regress first towards a refinement displacement y
s
i −
y
(s−1)
i
by applying a regressor on the sub image defined
by b
(s−1)
i
from previous stage (s − 1). Then, we estimate
new joint boxes b
s
i
:
Stage s: y
s
i ← y
(s−1)
i + N
−1
(ψi(N(x; b); θs); b)(6)
for b = b
(s−1)
i
b
s
i ← (y
s
i
, σdiam(y
s
), σdiam(y
s
)) (7)
We apply the cascade for a fixed number of stages S,
which is determined as explained in Sec. 4.1.
Training The network parameters θ1 are trained as
outlined in Sec. 3.1, Eq. (4). At subsequent stages
s ≥ 2, the training is done identically with one important difference. Each joint i from a training example (x, y) is normalized using a different bounding box
(y
(s−1)
i
, σdiam(y
(s−1)), σdiam(y
(s−1))) – the one centered at the prediction for the same joint obtained from previous stage – so that we condition the training of the stage
based on the model from previous stage.
Since deep learning methods have large capacity, we
augment the training data by using multiple normalizations
for each image and joint. Instead of using the prediction
from previous stage only, we generate simulated predictions. This is done by randomly displacing the ground truth
location for joint i by a vector sampled at random from a
2-dimensional Normal distribution N
(s−1)
i with mean and
variance equal to the mean and variance of the observed displacements (y
(s−1)
i − yi) across all examples in the training data. The full augmented training data can be defined
by first sampling an example and a joint from the original
data at uniform and then generating a simulated prediction
based on a sampled displacement δ from N
(s−1)
i
:
Ds
A = {(N(x; b), N(yi
; b))|
(x, yi) ∼ D, δ ∼ N (s−1)
i
,
b = (yi + δ, σdiam(y))}
The training objective for cascade stage s is done as in
Eq. (4) by taking extra care to use the correct normalization
for each joint:
θs = arg min
θ
X
(x,yi)∈Ds
A
||yi − ψi(x; θ)||2
2
(8)
4. Empirical Evaluation
4.1. Setup
Datasets There is a wide variety of benchmarks for human pose estimation. In this work we use datasets, which
have large number of training examples sufficient to train a
large model such as the proposed DNN, as well as are realistic and challenging.
The first dataset we use is Frames Labeled In Cinema
(FLIC), introduced by [19], which consists of 4000 training and 1000 test images obtained from popular Hollywood
movies. The images contain people in diverse poses and especially diverse clothing. For each labeled human, 10 upper
body joints are labeled.
The second dataset we use is Leeds Sports Dataset [12]
and its extension [13], which we will jointly denote by LSP.
Combined they contain 11000 training and 1000 testing images. These are images from sports activities and as such
are quite challenging in terms of appearance and especially
articulations. In addition, the majority of people have 150
pixel height which makes the pose estimation even more
challenging. In this dataset, for each person the full body is
labeled with total 14 joints.
For all of the above datasets, we define the diameter of a
pose y to be the distance between a shoulder and hip from
opposing sides and denote it by diam(y). It should be noted,
that the joints in all datasets are arranged in a tree kinematically mimicking the human body. This allows for a defini-
tion of a limb being a pair of neighboring joints in the pose
tree.
Metrics In order to be able to compare with published results we will use two widely accepted evaluation metrics.
Percentage of Correct Parts (PCP) measures detection rate
of limbs, where a limb is considered detected if the distance
between the two predicted joint locations and the true limb
joint locations is at most half of the limb length [5]. PCP
was the initially preferred metric for evaluation, however it
has the drawback of penalizing shorter limbs, such as lower
arms, which are usually harder to detect.
To address this drawback, recently detection rates of
joints are being reported using a different detection criterion – a joint is considered detected if the distance between
the predicted and the true joint is within a certain fraction of
the torso diameter. By varying this fraction, detection rates
are obtained for varying degrees of localization precision.
This metric alleviates the drawback of PCP since the detection criteria for all joints are based on the same distance
threshold. We refer to this metric as Percent of Detected
Joints (PDJ).
Experimental Details For all the experiments we use the
same network architecture. Inspired by [7], we use a body
detector on FLIC to obtain initially a rough estimate of the
human body bounding box. It is based on a face detector –
the detected face rectangle is enlarged by a fixed scaler. This
scaler is determined on the training data such that it contains
all labeled joints. This face-based body detector results in
a rough estimate, which however presents a good starting
point for our approach. For LSP we use the full image as
initial bounding box since the humans are relatively tightly
cropped by design.
Using a small held-out set of 50 images for both datasets
to determine the algorithm hyperparameters. To measure
optimality of the parameters we used average over PDJ at
0.2 across all joints. The scaler σ, which defines the size of
the refinement joint bounding box as a fraction of the pose
size, is determined as follows: for FLIC we chose σ = 1.0
after exploring values {0.8, 1.0, 1.2}, for LSP we use σ =
2.0 after trying {1.5, 1.7, 2.0, 2.3}. The number of cascade
stages S is determined by training stages until the algorithm
stopped improving on the held-out set. For both FLIC and
LSP we arrived at S = 3.
To improve generalization, for each cascade stage starting at s = 2 we augment the training data by sampling 40
randomly translated crop boxes for each joint as explained
in Sec. 3.2. Thus, for LSP with 14 joints and after mirroring the images and sampling the number training examples
is 11000 × 40 × 2 × 14 = 12M, which is essential for
training a large network as ours.
The presented algorithm allows for an efficient implementation. The running time is approx. 0.1s per image,
as measured on a 12 core CPU. This compares favorably
to other approaches, as some of the current state-of-art approaches have higher complexity: [19] runs in approx. 4s,
while [26] runs in 1.5s. The training complexity, however,
is higher. The initial stage was trained within 3 days on
approx. 100 workers, most of the final performance was
achieved after 12 hours though. Each refinement stage was
trained for 7 days since the amount of data was 40× larger
than the one for the initial stage due to the data augmentation in Sec. 3.2. Note that using more data led to increased
performance.
4.2. Results and Discussion
Comparisons We present comparative results to other approaches. We compare on LSP using PCP metric in Fig. 1.
We show results for the four most challenging limbs – lower
and upper arms and legs – as well as the average value
across these limbs for all compared algorithms. We clearly
outperform all other approaches, especially achieving better estimation for legs. For example, for upper legs we obtain 0.78 up from 0.74 for the next best performing method.
It is worth noting that while the other approaches exhibit
strengths for particular limbs, none of the other dataset consistently dominates across all limbs. In contrary, DeepPose
shows strong results for all challenging limbs.
Using the PDJ metric allows us to vary the threshold for
the distance between prediction and ground truth, which defines a detection. This threshold can be thought of as a
localization precision at which detection rates are plotted.
Thus one could compare approaches across different desired precisions. We present results on FLIC in Fig. 3 comparing against additional four methods as well is on LSP in
Fig. 4. For each dataset we train and test according the protocol for each dataset. Similarly to previous experiment we
outperform all five algorithms. Our gains are bigger in the
low precision domain, in the cases where we detect rough
pose without precisely localizing the joints. On FLIC, at
normalized distance 0.2 we obtain a an increase of detection
rates by 0.15 and 0.2 for elbow and wrists against the next
best performing method. On LSP, at normalized distance
0.5 we get an absolute increase of 0.1. At low precision
regime of normalized distance of 0.2 for LSP we show comparable performance for legs and slightly worse arms. This
can be attributed to the fact that the DNN-based approach
computes joint coordinates using 7 layers of transformation,
some of which contain max pooling.
Another observation is that our approach works well for
both appearance heavy movie data as well as string articulation such as the sports images in LSP.
Effects of cascade-based refinement A single DNNbased joint regressor gives rough joint location. However,
to obtain higher precision the subsequent stages of the cascade, which serve as a refinement of the initial prediction,
are of paramount importance. To see this, in Fig. 5 we
0 0.05 0.1 0.15 0.2
0.1
0.3
0.5
0.7
0.9
Elbows
Normalized distance to true joint
Detection rate
DeepPose
MODEC
Eichner et al.
Yang et al.
Sapp et al.
0 0.05 0.1 0.15 0.2
0.1
0.3
0.5
0.7
0.9
Wrists
Normalized distance to true joint
Detection rate
Figure 3. Percentage of detected joints (PDJ) on FLIC for two joints: elbow and wrist. We compare DeepPose, after two cascade stages,
with four other approaches.
Method Arm Leg Ave. Upper Lower Upper Lower
DeepPose-st1 0.5 0.27 0.74 0.65 0.54
DeepPose-st2 0.56 0.36 0.78 0.70 0.60
DeepPose-st3 0.56 0.38 0.77 0.71 0.61
Dantone et al. [2] 0.45 0.25 0.65 0.61 0.49
Tian et al. [24] 0.52 0.33 0.70 0.60 0.56
Johnson et al. [13] 0.54 0.38 0.75 0.66 0.58
Wang et al. [25] 0.565 0.37 0.76 0.68 0.59
Pishchulin [17] 0.49 0.32 0.74 0.70 0.56
Table 1. Percentage of Correct Parts (PCP) at 0.5 on LSP for DeepPose as well as five state-of-art approaches.
0 0.1 0.2 0.3 0.4 0.5
0.1
0.3
0.5
0.7
0.9
Arms
Normalized distance to true joint
Detection rate
DeepPose − wrists
DeepPose − elbows
Johnson et al. − wrists
Johnson et al. − elbows
0 0.1 0.2 0.3 0.4 0.5
0.1
0.3
0.5
0.7
0.9
Legs
Normalized distance to true joint
Detection rate
DeepPose − ankle
DeepPose − knee
Johnson et al. − ankle
Johnson et al. − knee
Figure 4. Percentage of detected joints (PDJ) on LSP for four
limbs for DeepPose and Johnson et al. [13] over an extended range
of distances to true joint: [0, 0.5] of the torso diameter. Results of
DeepPose are plotted with solid lines while all the results by [13]
are plotted in dashed lines. Results for the same joint from both
algorithms are colored with same color.
present the joint detections at different precisions for the initial prediction as well as two subsequent cascade stages. As
expected, we can see that the major gains of the refinement
procedure are at high-precision regime of at normalized distances of [0.15, 0.2]. Further, the major gains are achieved
after one stage of refinement. The reason being that subsequent stages end up using smaller sub-images around each
joint. And although the subsequent stages look at higher
resolution inputs, they have more limited context.
Examples of cases, where refinement helps, are visual0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
0.1
0.3
0.5
0.7
0.9
Wrists
Normalized distance to true joint
Detection rate
DeepPose − initial stage 1
DeepPose − stage 2
DeepPose − stage 3
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
0.1
0.3
0.5
0.7
0.9
Elbows
Normalized distance to true joint
Detection rate
Figure 5. Percent of detected joints (PDJ) on FLIC or the first three
stages of the DNN cascade. We present results over larger spectrum of normalized distances between prediction and ground truth.
ized in Fig. 6. The initial stage is usually successful at estimating a roughly correct pose, however, this pose is not
”snapped” to the correct one. For example, in row three the
pose has the right shape but incorrect scale. In the second
row, the predicted pose is translated north from the ideal
one. In most cases, the second stage of the cascade resolves
this snapping problem and better aligns the joints. In more
rare cases, such as in first row, further facade stages improve
on individual joints.
Cross-dataset Generalization To evaluate the generalization properties of our algorithm, we used the trained
models on LSP and FLIC on two related datasets. The fullbody model trained on LSP is tested on the test portion of
the Image Parse dataset [18] with results presented in Table 2. The ImageParse dataset is similar to LSP as it contains people doing sports, however it contains a lot of people from personal photo collections involved in other activities. Further, the upper-body model trained on FLIC was
applied on the whole Buffy dataset [7]. We can see that our
approach can retain state-of-art performance compared to
other approaches. This shows good generalization abilities.
Example poses To get a better idea of the performance of
our algorithm, we visualize a sample of estimated poses on
images from LSP in Fig. 8. We can see that our algorithm is
able to get correct pose for most of the joints under variety
of conditions: upside-down people (row 1, column 1), se-
Initial stage 1 stage 2 stage 3
Figure 6. Predicted poses in red and ground truth poses in green
for the first three stages of a cascade for three examples.
0 0.05 0.1 0.15 0.2
0.1
0.3
0.5
0.7
0.9
Elbows
Normalized distance to true joint
Detection rate
Eichner et al.
Yang et al.
Sapp et al.
MODEC
DeepPose
0 0.05 0.1 0.15 0.2
0.1
0.3
0.5
0.7
0.9
Wrists
Normalized distance to true joint
Detection rate
Figure 7. Percentage of detected joints (PDJ) on Buffy dataset
for two joints: elbow and wrist. The models have been trained on
FLIC. We compare DeepPose, after two cascade stages, with four
other approaches.
Method Arm Leg Ave. Upper Lower Upper Lower
DeepPose 0.8 0.75 0.71 0.5 0.69
Pishchulin [17] 0.80 0.70 0.59 037 0.62
Johnson et al. [13] 0.75 0.67 0.67 0.46 0.64
Yang et al. [26] 0.69 0.64 0.55 0.35 0.56
Table 2. Percentage of Correct Parts (PCP) at 0.5 on Image Parse
dataset for DeepPose as well as two state-of-art approaches on Image Parse dataset. Results obtained from [17].
vere foreshortening (row1, column 3), unusual poses (row
3, column 5), occluded limbs as the occluded arms in row
3, columns 2 and 6, unusual illumination conditions (row 3,
column 3). In most of the cases, when the estimated pose is
not precise, it still has a correct shape. For example, in the
last row some of the predicted limbs are not aligned with
the true locations, however the overall shape of the pose is
correct. A common failure mode is confusing left with right
side when the person was photographed from the back (row
6, column 6). Results on FLIC (see Fig. 9) are usually better
with occasional visible mistakes on lower arms.
5. Conclusion
We present, to our knowledge, the first application of
Deep Neural Networks (DNNs) to human pose estimation.
Our formulation of the problem as DNN-based regression to
joint coordinates and the presented cascade of such regressors has the advantage of capturing context and reasoning
about pose in a holistic manner. As a result, we are able to
achieve state-of-art or better results on several challenging
academic datasets.
Further, we show that using a generic convolutional neural network, which was originally designed for classification tasks, can be applied to the different task of localization. In future, we plan to investigate novel architectures
which could be potentially better tailored towards localization problems in general, and in pose estimation in particular.
Acknowledgements I would like to thank Luca Bertelli,
Ben Sapp and Tianli Yu for assistance with data and fruitful
discussions.
References
[1] M. Andriluka, S. Roth, and B. Schiele. Pictorial structures
revisited: People detection and articulated pose estimation.
In CVPR, 2009.
[2] M. Dantone, J. Gall, C. Leistner, and L. Van Gool. Human
pose estimation using body parts dependent joint regressors.
In CVPR, 2013.
[3] J. Duchi, E. Hazan, and Y. Singer. Adaptive subgradient
methods for online learning and stochastic optimization. In
COLT. ACL, 2010.
[4] M. Eichner and V. Ferrari. Better appearance models for
pictorial structures. 2009.
[5] M. Eichner, M. Marin-Jimenez, A. Zisserman, and V. Ferrari.
Articulated human pose estimation and search in (almost)
unconstrained still images. ETH Zurich, D-ITET, BIWI,
Technical Report No, 272, 2010.
[6] P. F. Felzenszwalb and D. P. Huttenlocher. Pictorial structures for object recognition. International Journal of Computer Vision, 61(1):55–79, 2005.
[7] V. Ferrari, M. Marin-Jimenez, and A. Zisserman. Progressive
search space reduction for human pose estimation. In CVPR,
2008.
[8] M. A. Fischler and R. A. Elschlager. The representation and
matching of pictorial structures. Computers, IEEE Transactions on, 100(1):67–92, 1973.
[9] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic
segmentation. In CVPR, 2014.
[10] G. Gkioxari, P. Arbelaez, L. Bourdev, and J. Malik. Articu- ´
lated pose estimation using discriminative armlet classifiers.
In CVPR, 2013.
[11] C. Ionescu, F. Li, and C. Sminchisescu. Latent structured
models for human pose estimation. In ICCV, 2011.
Figure 8. Visualization of pose results on images from LSP. Each pose is represented as a stick figure, inferred from predicted joints.
Different limbs in the same image are colored differently, same limb across different images has the same color.
Figure 9. Visualization of pose results on images from FLIC. Meaning of stick figures is the same as in Fig. 8 above.
[12] S. Johnson and M. Everingham. Clustered pose and nonlinear appearance models for human pose estimation. In BMVC,
2010.
[13] S. Johnson and M. Everingham. Learning effective human
pose estimation from inaccurate annotation. In CVPR, 2011.
[14] A. Krizhevsky, I. Sutskever, and G. Hinton. Imagenet classification with deep convolutional neural networks. In NIPS,
2012.
[15] G. Mori and J. Malik. Estimating human body configurations
using shape context matching. In ECCV, 2002.
[16] R. Nevatia and T. O. Binford. Description and recognition of
curved objects. Artificial Intelligence, 8(1):77–98, 1977.
[17] L. Pishchulin, M. Andriluka, P. Gehler, and B. Schiele. Poselet conditioned pictorial structures. In CVPR, 2013.
[18] D. Ramanan. Learning to parse images of articulated bodies.
In NIPS, 2006.
[19] B. Sapp and B. Taskar. Modec: Multimodal decomposable
models for human pose estimation. In CVPR, 2013.
[20] G. Shakhnarovich, P. Viola, and T. Darrell. Fast pose estimation with parameter-sensitive hashing. In CVPR, 2003.
[21] Y. Sun, X. Wang, and X. Tang. Deep convolutional network cascade for facial point detection. In Computer Vision
and Pattern Recognition (CVPR), 2013 IEEE Conference on,
pages 3476–3483. IEEE, 2013.
[22] C. Szegedy, A. Toshev, and D. Erhan. Object detection via
deep neural networks. In NIPS 26, 2013.
[23] G. W. Taylor, R. Fergus, G. Williams, I. Spiro, and C. Bregler. Pose-sensitive embedding by nonlinear nca regression.
In NIPS, 2010.
[24] Y. Tian, C. L. Zitnick, and S. G. Narasimhan. Exploring the
spatial hierarchy of mixture models for human pose estimation. In ECCV, 2012.
[25] F. Wang and Y. Li. Beyond physical connections: Tree models in human pose estimation. In CVPR, 2013.
[26] Y. Yang and D. Ramanan. Articulated pose estimation with
flexible mixtures-of-parts. In CVPR, 2011.
