Page 1

DeepCut: Joint Subset Partition and Labeling for Multi Person Pose Estimation

Leonid Pishchulin1, Eldar Insafutdinov1, Siyu Tang1, Bjoern Andres1,

Mykhaylo Andriluka1,3, Peter Gehler2, and Bernt Schiele1

1Max Planck Institute for Informatics, Germany

2Max Planck Institute for Intelligent Systems, Germany

3Stanford University, USA

Abstract

This paper considers the task of articulated human pose

estimation of multiple people in real world images. We pro-

pose an approach that jointly solves the tasks of detection

and pose estimation: it infers the number of persons in a

scene, identifies occluded body parts, and disambiguates

body parts between people in close proximity of each other.

This joint formulation is in contrast to previous strategies,

that address the problem by first detecting people and subse-

quently estimating their body pose. We propose a partition-

ing and labeling formulation of a set of body-part hypotheses

generated with CNN-based part detectors. Our formulation,

an instance of an integer linear program, implicitly performs

non-maximum suppression on the set of part candidates and

groups them to form configurations of body parts respect-

ing geometric and appearance constraints. Experiments on

four different datasets demonstrate state-of-the-art results

for both single person and multi person pose estimation1.

1. Introduction

Human body pose estimation methods have become in-

creasingly reliable. Powerful body part detectors [29] in

combination with tree-structured body models [30, 7] show

impressive results on diverse datasets [18, 3, 26]. These

benchmarks promote pose estimation of single pre-localized

persons but exclude scenes with multiple people. This prob-

lem definition has been a driver for progress, but also falls

short on representing a realistic sample of real-world images.

Many photographs contain multiple people of interest (see

Fig 1) and it is unclear whether single pose approaches gen-

eralize directly. We argue that the multi person case deserves

more attention since it is an important real-world task.

Key challenges inherent to multi person pose estimation

1Models and code available at http://pose.mpi-inf.mpg.de

(a)

(b)

(c)

Figure 1. Method overview: (a) initial detections (= part candidates)

and pairwise terms (graph) between all detections that (b) are jointly

clustered belonging to one person (one colored subgraph = one

person) and each part is labeled corresponding to its part class

(different colors and symbols correspond to different body parts);

(c) shows the predicted pose sticks.

are the partial visibility of some people, significant overlap

of bounding box regions of people, and the a-priori unknown

number of people in an image. The problem thus is to in-

fer the number of persons, assign part detections to person

instances while respecting geometric and appearance con-

straints. Most strategies use a two-stage inference process

[23, 15, 28] to first detect and then independently estimate

poses. This is unsuited for cases when people are in close

14929

proximity since they permit simultaneous assignment of the

same body-part candidates to multiple people hypotheses.

As a principled solution for multi person pose estimation

a model is proposed that jointly estimates poses of all people

present in an image by minimizing a joint objective. The

formulation is based on partitioning and labeling an initial

pool of body part candidates into subsets that correspond to

sets of mutually consistent body-part candidates and abide to

mutual consistency and exclusion constraints. The proposed

method has a number of appealing properties. (1) The for-

mulation is able to deal with an unknown number of people,

and also infers this number by linking part hypotheses. (2)

The formulation allows to either deactivate or merge part hy-

potheses in the initial set of part candidates hence effectively

performing non-maximum suppression (NMS). In contrast

to NMS performed on individual part candidates, the model

incorporates evidence from all other parts making the pro-

cess more reliable. (3) The problem is cast in the form of

an Integer Linear Program (ILP). Although the problem is

NP-hard, the ILP formulation facilitates the computation of

bounds and feasible solutions with a certified optimality gap.

This paper makes the following contributions. The main

contribution is the derivation of a joint detection and pose

estimation formulation cast as an integer linear program. Fur-

ther, two CNN variants are proposed to generate representa-

tive sets of body part candidates. These, combined with the

model, obtain state-of-the-art results for both single-person

and multi-person pose estimation on different datasets.

Related work. Most work on pose estimation targets the

single person case. Methods progressed from simple part

detectors and elaborate body models [25, 24, 16] to tree-

structured pictorial structures (PS) models with strong part

detectors [22, 34, 7, 26]. Impressive results are obtained pre-

dicting locations of parts with convolutional neural networks

(CNN) [31, 29]. While body models are not a necessary

component for effective part localization, constraints among

parts allow to assemble independent detections into body

configurations as demonstrated in [7] by combining CNN-

based body part detectors with a body model [34].

A popular approach to multi-person pose estimation is to

detect people first and then estimate body pose independently

[28, 23, 34, 15]. [34] proposes a flexible mixture-of-parts

model for detection and pose estimation. [34] obtains mul-

tiple pose hypotheses corresponding to different root part

positions and then performing non-maximum suppression.

[15] detects people using a flexible configuration of poselets

and the body pose is predicted as a weighted average of acti-

vated poselets. [23] detects people and then predicts poses

of each person using a PS model. [5] estimates poses of mul-

tiple people in 3D by constructing a shared space of 3D body

part hypotheses, but uses 2D person detections to establish

the number of people in the scene. These approaches are

limited to cases with people sufficiently far from each other

that do not have overlapping body parts.

Our work is closely related to [12, 21] who also propose

a joint objective to estimate poses of multiple people. [12]

proposes a multi-person PS model that explicitly models

depth ordering and person-person occlusions. Our formula-

tion is not limited by a number of occlusion states among

people. [21] proposes a joint model for pose estimation and

body segmentation coupling pose estimates of individuals

by image segmentation. [12, 21] uses a person detector to

generate initial hypotheses for the joint model. [21] resorts

to a greedy approach of adding one person hypothesis at a

time until the joint objective can be reduced, whereas our

formulation can be solved with a certified optimality gap.

In addition [21] relies on expensive labeling of body part

segmentation, which the proposed approach does not require.

Similarly to [8] we aim to distinguish between visible

and occluded body parts. [8] primarily focuse on the single-

person case and handles multi-person scenes akin to [34].

We consider the more difficult problem of full-body pose

estimation, whereas [12, 8] focus on upper-body poses and

consider a simplified case of people seen from the front.

Our work is related to early work on pose estimation

that also relies on integer linear programming to assemble

candidate body part hypotheses into valid configurations [16].

Their single person method employs a tree graph augmented

with weaker non-tree repulsive edges and expects the same

number of parts. In contrast, our novel formulation relies

on fully connected model to deal with unknown number of

people per image and body parts per person.

The Minimum Cost Multicut Problem [9, 11], known in

machine learning as correlation clustering [4], has been used

in computer vision for image segmentation [1, 2, 19, 35] but

has not been used before in the context of pose estimation.

It is known to be NP-hard [10].

2. Problem Formulation

In this section, the problem of estimating articulated poses

of an unknown number of people in an image is cast as an

optimization problem. The goal of this formulation is to state

three problems jointly: 1. The selection of a subset of body

parts from a set D of body part candidates, estimated from

an image as described in Section 4 and depicted as nodes of a

graph in Fig. 1(a). 2. The labeling of each selected body part

with one of C body part classes, e.g., “arm”, “leg”, “torso”,

as depicted in Fig. 1(c). 3. The partitioning of body parts

that belong to the same person, as depicted in Fig. 1(b).

2.1. Feasible Solutions

We encode labelings of the three problems jointly through

triples (x, y, z) of binary random variables with domains

x ∈ {0, 1}D×C,y ∈ {0, 1}(

D

2 ) and z ∈ {0, 1}(D

2 )×C2

.

Here, xdc = 1 indicates that body part candidate d is of

4930

class c, ydd′ = 1 indicates that the body part candidates d

and d′ belong to the same person, and zdd′cc′ are auxiliary

variables to relate x and y through zdd′cc′ = xdcxd′c′ ydd′ .

Thus, zdd′cc′

= 1 indicates that body part candidate d is

of class c (xdc = 1), body part candidate d′ is of class c′

(xd′c′ = 1), and body part candidates d and d′ belong to the

same person (ydd′ = 1).

In order to constrain the 01-labelings (x, y, z) to well-

defined articulated poses of one or more people, we impose

the linear inequalities (1)–(3) stated below. Here, the in-

equalities (1) guarantee that every body part is labeled with

at most one body part class. (If it is labeled with no body part

class, it is suppressed). The inequalities (2) guarantee that

distinct body parts d and d′ belong to the same person only if

neither d nor d′ is suppressed. The inequalities (3) guarantee,

for any three pairwise distinct body parts, d, d′ and d′′, if d

and d′ are the same person (as indicated by ydd′ = 1) and

d′ and d′′ are the same person (as indicated by yd′d′′ = 1),

then also d and d′′ are the same person (ydd′′ = 1), that is,

transitivity, cf. [9]. Finally, the inequalities (4) guarantee, for

any dd′ ∈ (D

2 ) and any cc′ ∈ C2 that zdd′cc′ = xdcxd′c′ ydd′ .

These constraints allow us to write an objective function as

a linear form in z that would otherwise be written as a cubic

form in x and y. We denote by XDC the set of all (x, y, z)

that satisfy all inequalities, i.e., the set of feasible solutions.

∀d ∈ D∀cc′ ∈ (C

2 ) : xdc + xdc′ ≤ 1

(1)

∀dd′ ∈ (D

2 ) : ydd′ ≤ ∑

c∈C

xdc

ydd′ ≤ ∑

c∈C

xd′c

(2)

∀dd′d′′ ∈ (D

3 ) : ydd′ + yd′d′′ − 1 ≤ ydd′′

(3)

∀dd′ ∈ (D

2 )∀cc′ ∈ C2 : xdc + xd′c′ + ydd′ − 2 ≤ zdd′cc′

zdd′cc′ ≤ xdc

zdd′cc′ ≤ xd′c′

zdd′cc′ ≤ ydd′

(4)

When at most one person is in an image, we further

constrain the feasible solutions to a well-defined pose of a

single person. This is achieved by an additional class of

inequalities which guarantee, for any two distinct body parts

that are not suppressed, that they must be clustered together:

∀dd′ ∈ (D

2 )∀cc′ ∈ C2 :

xdc + xd′c′ − 1 ≤ ydd′

(5)

2.2. Objective Function

For every pair (d, c) ∈ D × C, we will estimate a proba-

bility pdc ∈ [0, 1] of the body part d being of class c. In the

context of CRFs, these probabilities are called part unaries

and we will detail their estimation in Section 4.

For every dd′ ∈ (D

2 ) and every cc′ ∈ C2, we consider a

probability pdd′cc′ ∈ (0, 1) of the conditional probability of

d and d′ belonging to the same person, given that d and d′ are

body parts of classes c and c′, respectively. For c = c′, these

probabilities pdd′cc′ are the pairwise terms in a graphical

model of the human body. In contrast to the classic pictorial

structures model, our model allows for a fully connected

graph where each body part is connected to all other parts in

the entire set D by a pairwise term. For c = c′, pdd′cc′ is the

probability of the part candidates d and d′ representing the

same part of the same person. This facilitates clustering of

multiple part candidates of the same part of the same person

and a repulsive property that prevents nearby part candidates

of the same type to be associated to different people.

The optimization problem that we call the subset partition

and labeling problem is the ILP that minimizes over the set

of feasible solutions XDC:

min

(x,y,z)∈XDC

〈α, x〉 + 〈β,z〉,

(6)

where we used the short-hand notation

αdc := log

1 − pdc

pdc

(7)

βdd′cc′ := log

1 − pdd′cc′

pdd′cc′

(8)

〈α, x〉 := ∑

d∈D

∑

c∈C

αdc xdc

(9)

〈β,z〉 := ∑

dd′∈(D

2 )

∑

c,c′∈C

βdd′cc′ zdd′cc′

.

(10)

The objective (6)–(10) is the MAP estimate of a prob-

ability measure of joint detections x and clusterings y, z

of body parts, where prior probabilities pdc and pdd′cc′ are

estimated independently from data, and the likelihood is a

positive constant if (x, y, z) satisfies (1)–(4), and is 0, other-

wise. The exact form (6)–(10) is obtained when minimizing

the negative logarithm of this probability measure.

2.3. Optimization

In order to obtain feasible solutions of the ILP (6) with

guaranteed bounds, we separate the inequalities (1)–(5) in

the branch-and-cut loop of the state-of-the-art ILP solver

Gurobi. More precisely, we solve a sequence of relaxations

of the problem (6), starting with the (trivial) unconstrained

problem. Each problem is solved using the cuts proposed

by Gurobi. Once an integer feasible solution is found, we

identify violated inequalities (1)–(5), if any, by breadth-first-

search, add these to the constraint pool and re-solve the

tightened relaxation. Once an integer solution satisfying

all inequalities is found, together with a lower bound that

certifies an optimality gap below 1%, we terminate.

3. Pairwise Probabilities

Here we describe the estimation of the pairwise terms.

We define pairwise features fdd′

for the variable zdd′cc′

4931

(Sec. 2). Each part detection d includes the probabilities

fpdc (Sec. 4.4), its location (xd,yd), scale hd and bound-

ing box Bd coordinates. Given two detections d and d′,

and the corresponding features (fpdc ,xd,yd,hd,Bd) and

(fpd′c ,xd′ ,yd′ ,hd′ ,Bd′ ), we define two sets of auxiliary

variables for zdd′cc′ , one set for c = c′ (same body part class

clustering) and one for c = c′ (across two body part classes

labeling). These features capture the proximity, kinematic

relation and appearance similarity between body parts.

The same body part class (c = c′). Two detections de-

noting the same body part of the same person should be

in close proximity to each other. We introduce the fol-

lowing auxiliary variables that capture the spatial relations:

∆x = |xd −xd′ |/¯h, ∆y = |yd −yd′ |/¯h, ∆h = |hd −hd′ |/¯h,

IOUnion, IOMin, IOMax. The latter three are intersec-

tions over union/minimum/maximum of the two detection

boxes, respectively, and ¯h = (hd + hd′ )/2.

Non-linear Mapping. We augment the feature repre-

sentation by appending quadratic and exponential terms.

The final pairwise feature fdd′

for the variable zdd′cc is

(∆x, ∆y, ∆h,IOUnion,IOMin,IOMax, (∆x)

2

,

..., (IOMax)

2

, exp (−∆x),..., exp (−IOMax)).

Two different body part classes (c = c′). We encode the

kinematic body constraints into the pairwise feature by in-

troducing auxiliary variables Sdd′ and Rdd′ , where Sdd′ and

Rdd′ are the Euclidean distance and the angle between two

detections, respectively. To capture the joint distribution of

Sdd′ and Rdd′ , instead of using Sdd′ and Rdd′ directly, we

employ the posterior probability p(zdd′cc′ = 1|Sdd′ ,Rdd′ )

as pairwise feature for zdd′cc′ to encode the geometric rela-

tions between the body part class c and c′. More specifically,

assuming the prior probability p(zdd′cc′ = 1) = p(zdd′cc′ =

0) = 0.5, the posterior probability of detection d and d′ have

the body part label c and c′, namely zdd′cc′ = 1, is

p(zdd′cc′ = 1|Sdd′ ,Rdd′ )

=

p(Sdd′ ,Rdd′ |zdd′cc′ = 1)

p(Sdd′ ,Rdd′ |zdd′cc′ = 1) + p(Sdd′ ,Rdd′ |zdd′cc′ = 0)

,

where p(Sdd′ ,Rdd′ |zdd′cc′ = 1) is obtained by conducting

a normalized 2D histogram of Sdd′ and Rdd′ from posi-

tive training examples, analogous to the negative likelihood

p(Sdd′ ,Rdd′ |zdd′cc′ = 0). In Sec. 5.1 we also experiment

with encoding the appearance into the pairwise feature by

concatenating the feature fpdc from d and fpd′c from d′, as

fpdc is the output of the CNN-based part detectors. The final

pairwise feature is (p(zdd′cc′ = 1|Sdd′ ,Rdd′ ),fpdc ,fpd′c ).

3.1. Probability Estimation

The coefficients α and β of the objective function (Eq. 6)

are defined by the probability ratio in the log space (Eq. 7 and

Eq. 8). Here we describe the estimation of the corresponding

probability density: (1) For every pair of detection and part

classes, namely for any (d, c) ∈ D × C, we estimate a

probability pdc ∈ (0, 1) of the detection d being a body

part of class c. (2) For every combination of two distinct

detections and two body part classes, namely for any dd′ ∈

(D

2 ) and any cc′ ∈ C2, we estimate a probability pdd′cc′ ∈

(0, 1) of d and d′ belonging to the same person, meanwhile

d and d′ are body parts of classes c and c′, respectively.

Learning. Given the features fdd′ and a Gaussian prior

p(θcc′ ) = N(0,σ2) on the parameters, logistic model is

p(zdd′cc′ = 1|fdd′ ,θcc′ ) =

1

1 + exp(−〈θcc′ ,fdd′ 〉)

. (11)

(|C| × (|C| + 1))/2 parameters are estimated using ML.

Inference Given two detections d and d′, the coefficients

αdc for xdc and αd′c for xd′c are obtained by Eq. 7, the

coefficient βdd′cc′ for zdd′cc′ has the form

βdd′cc′ = log

1 − pdd′cc′

pdd′cc′

= −〈fdd′ ,θcc′ 〉.

(12)

Model parameters θcc′ are learned using logistic regression.

4. Body Part Detectors

We first introduce our deep learning-based part detection

models and then evaluate them on two prominent bench-

marks thereby significantly outperforming state of the art.

4.1. Adapted Fast R-CNN (AFR-CNN)

To obtain strong part detectors we adapt Fast R-

CNN [14]. FR-CNN takes as input an image and set of

class-independent region proposals [32] and outputs the soft-

max probabilities over all classes and refined bounding boxes.

To adapt FR-CNN for part detection we alter it in two ways:

1) proposal generation and 2) detection region size. The

adapted version is called AFR-CNN throughout the paper.

Detection proposals. Generating object proposals is essen-

tial for FR-CNN, meanwhile detecting body parts is chal-

lenging due to their small size and high intra-class variability.

We use DPM-based part detectors [22] for proposal gener-

ation. We collect K top-scoring detections by each part

detector in a common pool of N part-independent proposals

and use these proposals as input to AFR-CNN. N is 2, 000

in case of single and 20, 000 in case of multiple people.

Larger context. Increasing the size of DPM detections

by upscaling every bounding box by a fixed factor allows

to capture more context around each part. In Sec. 4.3 we

evaluate the influence of upscaling and show that using larger

context around parts is crucial for best performance.

Details. Following standard FR-CNN training procedure Im-

ageNet models are finetuned on pose estimation task. Center

of a predicted bounding box is used for body part location

prediction. See supplemental for detailed parameter analysis.

4932

Setting

Head Sho Elb Wri Hip Knee Ank PCK AUC

oracle2,000

98.8 98.8 97.4 96.4 97.4 98.3 97.7 97.8 84.0

DPMscale1

48.8 25.1 14.4 10.2 13.6 21.8 27.1 23.0 13.6

AlexNetscale1

82.2 67.0 49.6 45.4 53.1 52.9 48.2 56.9 35.9

AlexNetscale4

85.7 74.4 61.3 53.2 64.1 63.1 53.8 65.1 39.0

+optimalparams

88.1 79.3 68.9 62.6 73.5 69.3 64.7 72.4 44.6

VGGscale4optimalparams 91.0 84.2 74.6 67.7 77.4 77.3 72.8 77.9 50.0

+ finetune LSP

95.4 86.5 77.8 74.0 84.5 78.8 82.6 82.8 57.0

Table 1. Unary only performance (PCK) of AFR-CNN on the LSP

(Person-Centric) dataset. AFR-CNN is finetuned from ImageNet to

MPII (lines 3-6), and then finetuned to LSP (line 7).

4.2. Dense Architecture (Dense-CNN)

Using proposals for body part detection may be sub-

optimal. We thus develop a fully convolutional architecture

for computing part probability scoremaps.

Stride. We build on VGG [27]. Fully convolutional VGG

has stride of 32 px – too coarse for precise part localization.

We thus use hole algorithm [6] to reduce the stride to 8 px.

Scale. Selecting image scale is crucial. We found that scal-

ing to a standing height of 340 px performs best: VGG

receptive field sees entire body to disambiguate body parts.

Loss function. We start with a softmax loss that outputs

probabilities for each body part and background. The down-

side is inability to assign probabilities above 0.5 to several

close-by body parts. We thus re-formulate the detection as

multi-label classification, where at each location a separate

set of probability distributions is estimated for each part. We

use sigmoid activation function on the output neurons and

cross entropy loss. We found this loss to perform better than

softmax and converge much faster compared to MSE [30].

Target training scoremap for each joint is constructed by

assigning a positive label 1 at each location within 15 px to

the ground truth, and negative label 0 otherwise.

Location refinement. In order to improve location preci-

sion we follow [14]: we add a location refinement FC layer

after the FC7 and use the relative offsets (∆x, ∆y) from a

scoremap location to the ground truth as targets.

Regression to other parts. Similar to location refinement

we add an extra term to the objective function where for each

part we regress onto all other part locations. We found this

auxiliary task to improve the performance (c.f. Sec. 4.3).

Training. We follow best practices and use SGD for CNN

training. In each iteration we forward-pass a single image.

After FC6 we select all positive and random negative sam-

ples to keep the pos/neg ratio as 25%/75%. We finetune

VGG from Imagenet model to pose estimation task and use

training data augmentation. We train for 430k iterations with

the following learning rates (lr): 10k at lr=0.001, 180k at

lr=0.002, 120k at lr=0.0002 and 120k at lr=0.0001. Pre-

training at smaller lr prevents the gradients from diverging.

Setting

Head Sho Elb Wri Hip Knee Ank PCK AUC

MPIIsoftmax

91.5 85.3 78.0 72.4 81.7 80.7 75.7 80.8 51.9

+LSPET

94.6 86.8 79.9 75.4 83.5 82.8 77.9 83.0 54.7

+sigmoid

93.5 87.2 81.0 77.0 85.5 83.3 79.3 83.8 55.6

+locationrefinement 95.0 88.4 81.5 76.4 88.0 83.3 80.8 84.8 61.5

+auxiliarytask

95.1 89.6 82.8 78.9 89.0 85.9 81.2 86.1 61.6

+ finetune LSP

97.2 90.8 83.0 79.3 90.6 85.6 83.1 87.1 63.6

Table 2. Unary only performance (PCK) of Dense-CNN VGG on

LSP (PC) dataset. Dense-CNN is finetuned from ImageNet to MPII

(line 1), to MPII+LSPET (lines 2-5), and finally to LSP (line 6).

4.3. Evaluation of Part Detectors

Datasets. We train and evaluate on three public benchmarks:

“Leeds Sports Poses” (LSP) [17] (person-centric (PC)), “LSP

Extended” (LSPET) [18]2, and “MPII Human Pose” (“Sin-

gle Person”) [3]. The MPII training set (19185 people) is

used as default. In some cases LSP training and LSPET are

added to MPII (marked as MPII+LSPET in the experiments).

Evaluation measures. We use the standard “PCK” met-

ric [26, 31, 30] and evaluation scripts available on the web

page of [3]. In addition, we report “Area under Curve”

(AUC) computed for the entire range of PCK thresholds.

AFR-CNN. Evaluation of AFR-CNN on LSP is shown in

Tab. 1. Oracle selecting per part the closest from 2, 000 pro-

posals achieves 97.8% PCK, as proposals cover majority of

the ground truth locations. Choosing a single proposal per

part using DPM score achieves 23.0% PCK – not surprising

given the difficulty of the body part detection problem. Re-

scoring the proposals using AFR-CNN with AlexNet [20]

dramatically improves the performance to 56.9% PCK, as

CNN learns richer image representations. Extending the

regions by 4x (1x ≈ head size) achieves 65.1% PCK, as it

incorporates more context including the information about

symmetric parts and allows to implicitly encode higher-order

part relations. Using data augmentation and slightly tun-

ing training parameters improves the performance to 72.4%

PCK. We refer to the supplementary material for detailed

analysis. Deeper VGG architecture improves over smaller

AlexNet reaching 77.9% PCK. All results so far are achieved

by finetuning the ImageNet models on MPII. Further fine-

tuning to LSP leads to remarkable 82.8% PCK: CNN learns

LSP-specific image representations. Strong increase in AUC

(57.0 vs. 50%) is due to improvements for smaller PCK

thresholds. Using no bounding box regression leads to per-

formance drop (81.3% PCK, 53.2% AUC): location refine-

ment is crucial for better localization. Overall AFR-CNN

obtains very good results on LSP by far outperforming the

state of the art (c.f. Tab. 3, rows 7 − 9). Evaluation on MPII

shows competitive performance (Tab. 4, row 1).

Dense-CNN. The results are in Tab. 2. Training with VGG

on MPII with softmax loss achieves 80.8% PCK thereby

2To reduce labeling noise we re-annotated original high-resolution im-

ages and make the data available at http://datasets.d2.mpi-inf.

mpg.de/hr-lspet/hr-lspet.zip

4933

Normalized distance

0

0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2

Detection rate, %

0

10

20

30

40

50

60

70

80

90

100

PCK total, LSP PC

AFR-CNN

DeepCut SP AFR-CNN

Dense-CNN

DeepCut SP Dense-CNN

Tompson et al., NIPS'14

Chen&Yuille, NIPS'14

Fan et al., CVPR'15

Normalized distance

0

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5

Detection rate, %

0

10

20

30

40

50

60

70

80

90

100

PCKh total, MPII

AFR-CNN

DeepCut SP AFR-CNN

Dense-CNN

DeepCut SP Dense-CNN

Tompson et al., NIPS'14

Tompson et al., CVPR'15

(a) LSP (PC)

(b) MPII Single Person

Figure 2. Pose estimation results over all PCK thresholds.

outperforming AFR-CNN (c.f. Tab. 1, row 6). This shows

the advantages of fully convolutional training and evalua-

tion. Expectedly, training on larger MPII+LSPET dataset

improves the results (83.0 vs. 80.8% PCK). Using cross-

entropy loss with sigmoid activations improves the results to

83.8% PCK, as it better models the appearance of close-by

parts. Location refinement improves localization accuracy

(84.8% PCK), which becomes more clear when analyzing

AUC (61.5 vs. 55.6%). Interestingly, regressing to other

parts further improves PCK to 86.1% showing a value of

training with the auxiliary task. Finally, finetuning to LSP

achieves the best result of 87.1% PCK, which is significantly

higher than the best published results (c.f. Tab. 3, rows 7−9).

Unary-only evaluation on MPII reveals slightly higher AUC

results compared to the state of the art (Tab. 4, row 3 − 4).

4.4. Using Detections in DeepCut Models

The SPLP problem is NP-hard, to solve instances of it

efficiently we select a subset of representative detections

from the entire set produced by a model. In our experiments

we use |D| = 100 as default detection set size. In case of the

AFR-CNN we directly use the softmax output as unary prob-

abilities: fpdc = (pd1,...,pdc), where pdc is the probability

of the detection d being the part class c. For Dense-CNN

detection model we use the sigmoid detection unary scores.

5. DeepCut Results

The aim of this paper is to tackle the multi person case.

To that end, we evaluate the proposed DeepCut models on

four diverse benchmarks. We confirm that both single person

(SP) and multi person (MP) variants (Sec. 2) are effective

on standard SP pose estimation datasets [17, 3]. Then, we

demonstrate superior performance of DeepCut MP on the

multi person pose estimation task.

5.1. Single Person Pose Estimation

We now evaluate single person (SP) and more general

multi person (MP) DeepCut models on LSP and MPII SP

benchmarks described in Sec. 4. Since this evaluation setting

implicitly relies on the knowledge that all parts are present

in the image we always output the full number of parts.

Results on LSP. We report per-part PCK results (Tab. 3)

and results for a variable distance threshold (Fig. 2 (a)).

Setting

Head Sho Elb Wri Hip Knee Ank PCK AUC

AFR-CNN (unary)

95.4 86.5 77.8 74.0 84.5 82.6 78.8 82.8 57.0

+ DeepCut SP

95.4 86.7 78.3 74.0 84.3 82.9 79.2 83.0 58.4

+appearancepairwise 95.4 87.2 78.6 73.7 84.7 82.8 78.8 83.0 58.5

+ DeepCut MP

95.2 86.7 78.2 73.5 84.6 82.8 79.0 82.9 58.0

Dense-CNN (unary)

97.2 90.8 83.0 79.3 90.6 85.6 83.1 87.1 63.6

+ DeepCut SP

97.0 91.0 83.8 78.1 91.0 86.7 82.0 87.1 63.5

+ DeepCut MP

96.2 91.2 83.3 77.6 91.3 87.0 80.4 86.7 62.6

Tompsonetal.[30]

90.6 79.2 67.9 63.4 69.5 71.0 64.2 72.3 47.3

Chen&Yuille[7]

91.8 78.2 71.8 65.5 73.3 70.2 63.4 73.4 40.1

Fan et al. [33]∗

92.4 75.2 65.3 64.0 75.7 68.3 70.4 73.0 43.2

∗ re-evaluated using the standard protocol, for details see project page of [33]

Table 3. Pose estimation results (PCK) on LSP (PC) dataset.

DeepCut SP AFR-CNN model using 100 detections im-

proves over unary only (83.0 vs. 82.8% PCK, 58.4 vs.

57% AUC), as pairwise connections filter out some of the

high-scoring detections on the background. The improve-

ment is clear in Fig. 2 (a) for smaller thresholds. Using

part appearance scores in addition to geometrical features in

c = c′ pairwise terms only slightly improves AUC, as the

appearance of neighboring parts is mostly captured by a rela-

tively large region centered at each part. The performance of

DeepCut MP AFR-CNN matches the SP and improves over

AFR-CNN alone: DeepCut MP correctly handles the SP case.

Performance of DeepCut SP Dense-CNN is almost identical

to unary only, unlike the results for AFR-CNN. Dense-CNN

performance is noticeably higher compared to AFR-CNN,

and “easy” cases that could have been corrected by a spatial

model are resolved by stronger part detectors alone.

Comparison to the state of the art (LSP). Tab. 3 compares

results of DeepCut models to other deep learning methods

specifically designed for single person pose estimation. All

DeepCuts significantly outperform the state of the art, with

DeepCut SP Dense-CNN model improving by 13.7% PCK

over the best known result [7]. The improvement is even

more dramatic for lower thresholds (Fig. 2 (a)): for PCK

@ 0.1 the best model improves by 19.9% over Tompson et

al. [30], by 26.7% over Fan et al. [33], and by 32.4% PCK

over Chen&Yuille [7]. The latter is interesting, as [7] use a

stronger spatial model that predicts the pairwise conditioned

on the CNN features, whereas DeepCuts use geometric-only

pairwise connectivity. Including body part orientation infor-

mation into DeepCuts should further improve the results.

Results on MPII Single Person. Results are shown in

Tab. 4 and Fig. 2 (b). DeepCut SP AFR-CNN noticeably

improves over AFR-CNN alone (79.8 vs. 78.8% PCK, 51.1

vs. 49.0% AUC). The improvement is stronger for smaller

thresholds (c.f. Fig. 2), as spatial model improves part local-

ization. Dense-CNN alone trained on MPII outperforms

AFR-CNN (81.6 vs. 78.8% PCK), which shows the ad-

vantages of dense training and evaluation. As expected,

Dense-CNN performs slightly better when trained on the

larger MPII+LSPET. Finally, DeepCut Dense-CNN SP is

slightly better than Dense-CNN alone leading to the best

4934

Setting

Head Sho Elb Wri Hip Knee Ank PCKh AUC

AFR-CNN (unary) 91.5 89.7 80.5 74.4 76.9 69.6 63.1 78.8 49.0

+ DeepCut SP

92.3 90.6 81.7 74.9 79.2 70.4 63.0 79.8 51.1

Dense-CNN (unary) 93.5 88.6 82.2 77.1 81.7 74.4 68.9 81.6 56.0

+LSPET

94.0 89.4 82.3 77.5 82.0 74.4 68.7 81.9 56.5

+DeepCut SP

94.1 90.2 83.4 77.3 82.6 75.7 68.6 82.4 56.5

Tompson et al. [30] 95.8 90.3 80.5 74.3 77.6 69.7 62.8 79.6 51.8

Tompson et al. [29] 96.1 91.9 83.9 77.8 80.9 72.3 64.8 82.0 54.9

Table 4. Pose estimation results (PCKh) on MPII Single Person.

result on MPII dataset (82.4% PCK).

Comparison to the state of the art (MPII). We com-

pare the performance of DeepCut models to the best

deep learning approaches from the literature [30, 29]3.

DeepCut SP Dense-CNN outperforms both [30, 29] (82.4

vs 79.6 and 82.0% PCK, respectively). Similar to them

DeepCuts rely on dense training and evaluation of part de-

tectors, but unlike them use single size receptive field and

do not include multi-resolution context information. Also,

appearance and spatial components of DeepCuts are trained

piece-wise, unlike [30]. We observe that performance dif-

ferences are higher for smaller thresholds (c.f. Fig. 2 (b)).

This is remarkable, as a much simpler strategy for location

refinement is used compared to [29]. Using multi-resolution

filters and joint training should improve the performance.

5.2. Multi Person Pose Estimation

We now evaluate DeepCut MP models on the challenging

task of MP pose estimation with an unknown number of

people per image and visible body parts per person.

Datasets. For evaluation we use two public MP benchmarks:

“We Are Family” (WAF) [12] with 350 training and 175 test-

ing group shots of people; “MPII Human Pose” (“Multi-

Person”) [3] consisting of 3844 training and 1758 testing

groups of multiple interacting individuals in highly articu-

lated poses with variable number of parts. On MPII, we use

a subset of 288 testing images for evaluation. We first pre-

finetune both AFR-CNN and Dense-CNN from ImageNet to

MPII and MPII+LSPET, respectively, and further finetune

each model to WAF and MPII Multi-Person. For WAF, we

re-train the spatial model on WAF training set.

WAF evaluation measure. Approaches are evaluated using

the official toolkit [12], thus results are directly comparable

to prior work. The toolkit implements occlusion-aware “Per-

centage of Correct Parts (mPCP)” metric. In addition, we

report “Accuracy of Occlusion Prediction (AOP)” [8].

MPII Multi-Person evaluation measure. PCK metric is

suitable for SP pose estimation with known number of parts

and does not penalize for false positives that are not a part

of the ground truth. Thus, for MP pose estimation we

use “Mean Average Precision (mAP)” measure, similar to

[28, 34]. In contrast to [28, 34] evaluating the detection

3[30] was re-trained and evaluated on MPII dataset by the authors.

Setting

Head U Arms L Arms Torso mPCP AOP

AFR-CNN det ROI

69.8

46.0

36.7

83.7

53.1 73.9

DeepCut MP AFR-CNN

99.0

79.5

74.3

87.1

82.2 85.6

Dense-CNN det ROI

76.0

46.0

40.2

83.7

55.3 73.8

DeepCut MP Dense-CNN 99.3

81.5

79.5

87.1

84.7 86.5

Ghiasi et. al. [13]

-

-

-

-

63.6 74.0

Eichner&Ferrari [12]

97.6

68.2

48.1

86.1

69.4 80.0

Chen&Yuille [8]

98.5

77.2

71.3

88.5

80.7 84.9

Table 5. Pose estimation results (mPCP) on WAF dataset.

of any part instance in the image disrespecting inconsis-

tent pose predictions, we evaluate consistent part configura-

tions. First, multiple body pose predictions are generated and

then assigned to the ground truth (GT) based on the highest

PCKh [3]. Only single pose can be assigned to GT. Unas-

signed predictions are counted as false positives. Finally, AP

for each body part is computed and mAP is reported.

Baselines. To assess the performance of AFR-CNN and

Dense-CNN we follow a traditional route from the literature

based on two stage approach: first a set of regions of inter-

est (ROI) is generated and then the SP pose estimation is

performed in the ROIs. This corresponds to unary only per-

formance. ROI are either based on a ground truth (GT ROI)

or on the people detector output (det ROI).

Results on WAF. Results are shown in Tab. 5. det ROI is

obtained by extending provided upper body detection boxes.

AFR-CNN det ROI achieves 57.6% mPCP and 73.9% AOP.

DeepCut MP AFR-CNN significantly improves over AFR-

CNN det ROI achieving 82.2% mPCP. This improvement

is stronger compared to LSP and MPII due to several rea-

sons. First, mPCP requires consistent prediction of body

sticks as opposite to body joints, and including spatial model

enforces consistency. Second, mPCP metric is occlusion-

aware. DeepCuts can deactivate detections for the occluded

parts thus effectively reasoning about occlusion. This is

supported by strong increase in AOP (85.6 vs. 73.9%). Re-

sults by DeepCut MP Dense-CNN follow the same tendency

achieving the best performance of 84.7% mPCP and 86.5%

AOP. Both increase in mPCP and AOP show the advantages

of DeepCuts over traditional det ROI approaches.

Tab. 5 shows that DeepCuts outperform all prior methods.

Deep learning method [8] is outperformed both for mPCP

(84.7 vs. 80.7%) and AOP (86.5 vs. 84.9%) measures. This

is remarkable, as DeepCuts reason about part interactions

across several people, whereas [8] primarily focuses on the

single-person case and handles multi-person scenes akin to

[34]. In contrast to [8], DeepCuts are not limited by the num-

ber of possible occlusion patterns and cover person-person

occlusions and other types as truncation and occlusion by

objects in one formulation. DeepCuts significantly outper-

form [12] while being more general: unlike [12] DeepCuts

do not require person detector and not limited by a number

of occlusion states among people.

Qualitative comparison to [8] is provided in Fig. 3.

4935

det

R

OI

1

2

3

4

5

6

1

2

3

4

5

6

1

2

3

4

5

6

1

2

3

4

5

6

1

2

3

4

5

6

1

2 3

45

6

1

2

3

4

5

6

12

3

4

5

6

1

2

3

4

5

6

12

3

4

5

6

12 3

4

5

6

1

2

3

4

5

6

1

2

3

4

5

6

1

2

3

45

6

1

2

3

4

5

6

1

2

3

4

5

6

1

23

45

6

1

2 3

4

5

6

1

2

3

4

5

6

1

23

4

5

6

1

2

3

4 5

6

1

23

4

5

6

DeepCut

MP

1

2

3

5

6

1

2

3

4

5

6

1

2

3

4

6

1

2

3

4

5

6

1 3

5

6

1

2

3

4 5

6

1

2

3

4

5

6

1

2

4

6

1

2

3

4

6

1

2

3

4

6

1

2

4

6

1

2

3

4

5

6

1

2

3

4

5

6

1

2

3

4

5

6

1

2

4

6

1

2

3

4

5

6

1

3

5

6

1

2

3

4

5

6

1

2

3

4

5

6

1

2

4

6

1

6

1

3

5

6

Chen&Y

uille

[

8

]

6

1

2

3

4

5

6

6

1

2

3

4

5

6

1

3

5

6

1

2

3

4

5

6

1

2

3

4

5

6

1

2

3

4

6

1

2

3

4

5

6

1

2

3

4

5

6

1

2

3

4

5

6

1

2

4

6

1

2

6

1

2

3

4

5

6

1

2

3

4

6

1

2

3

4

5

6

1

2

3

4

5

6

1

3

5

6

1

2

3

4

5

6

1

2

4

6

12

3

4

6

1

3

5

6

Figure 3. Qualitative comparison of our joint formulation DeepCut MP Dense-CNN (middle) to the traditional two-stage approach

Dense-CNN det ROI (top) and the approach of Chen&Yuille [8] (bottom) on WAF dataset. In contrast to det ROI, DeepCut MP is able

to disambiguate multiple and potentially overlapping persons and correctly assemble independent detections into plausible body part

configurations. In contrast to [8], DeepCut MP can better predict occlusions (image 2 person 1 − 4 from the left, top row; image 4 person 1,

4; image 5, person 2) and better cope with strong articulations and foreshortenings (image 1, person 1, 3; image 2 person 1 bottom row;

image 3, person 1-2). See supplementary material for more examples.

Results on MPII Multi-Person. Obtaining a strong detec-

tor of highly articulated people having strong occlusions and

truncations is difficult. We employ a neck detector as a per-

son detector as it turned out to be the most reliable part. Full

body bounding box is created around a neck detection and

used as det ROI. GT ROIs were provided by the authors [3].

As the MP approach [8] is not public, we compare to SP

state-of-the-art method [7] applied to GT ROI image crops.

Results are shown in Tab. 6. DeepCut MP AFR-CNN

improves over AFR-CNN det ROI by 4.3% achieving 51.4%

AP. The largest differences are observed for the ankle, knee,

elbow and wrist, as those parts benefit more from the con-

nections to other parts. DeepCut MP UB AFR-CNN using

upper body parts only slightly improves over the full body

model when compared on common parts (60.5 vs 58.2% AP).

Similar tendencies are observed for Dense-CNNs, though im-

provements of MP UB over MP are more significant.

All DeepCuts outperform Chen&Yuille SP GT ROI, par-

tially due to stronger part detectors compared to [7] (c.f.

Tab. 3). Another reason is that Chen&Yuille SP GT ROI does

not model body part occlusion and truncation always pre-

dicting the full set of parts, which is penalized by the AP

measure. In contrast, our formulation allows to deactivate the

part hypothesis in the initial set of part candidates thus effec-

tively performing non-maximum suppression. In DeepCuts

part hypotheses are suppressed based on the evidence from

all other body parts making this process more reliable.

Setting

Head Sho Elb Wri Hip Knee Ank UBody FBody

AFR-CNN det ROI

71.1 65.8 49.8 34.0 47.7 36.6 20.6 55.2

47.1

AFR-CNN MP

71.8 67.8 54.9 38.1 52.0 41.2 30.4 58.2

51.4

AFR-CNN MP UB

75.2 71.0 56.4 39.6 -

-

-

60.5

-

Dense-CNN det ROI

77.2 71.8 55.9 42.1 53.8 39.9 27.4 61.8

53.2

Dense-CNN MP

73.4 71.8 57.9 39.9 56.7 44.0 32.0 60.7

54.1

Dense-CNN MP UB

81.5 77.3 65.8 50.0 -

-

-

68.7

-

AFR-CNN GT ROI

73.2 66.5 54.6 42.3 50.1 44.3 37.8 59.1

53.1

Dense-CNN GT ROI

78.1 74.1 62.2 52.0 56.9 48.7 46.1 66.6

60.2

Chen&Yuille SP GT ROI 65.0 34.2 22.0 15.7 19.2 15.8 14.2 34.2

27.1

Table 6. Pose estimation results (AP) on MPII Multi-Person.

6. Conclusion

Articulated pose estimation of multiple people in uncon-

trolled real world images is challenging but of real world

interest. In this work, we proposed a new formulation as a

joint subset partitioning and labeling problem (SPLP). Dif-

ferent to previous two-stage strategies that separate the de-

tection and pose estimation steps, the SPLP model jointly

infers the number of people, their poses, spatial proxim-

ity, and part level occlusions. Empirical results on four

diverse and challenging datasets show significant improve-

ments over all previous methods not only for the multi per-

son, but also for the single person pose estimation problem.

On multi person WAF dataset we improve by 30% PCP over

the traditional two-stage approach. This shows that a joint

formulation is crucial to disambiguate multiple and poten-

tially overlapping persons. Models and code available at

http://pose.mpi-inf.mpg.de.

4936

References

[1] A. Alush and J. Goldberger. Ensemble segmentation using

efficient integer linear programming. TPAMI, 34(10):1966–

1977, 2012. 2

[2] B. Andres, J. H. Kappes, T. Beier, U. Köthe, and F. A. Ham-

precht. Probabilistic image segmentation with closedness

constraints. In ICCV, 2011. 2

[3] M. Andriluka, L. Pishchulin, P. Gehler, and B. Schiele. 2d

human pose estimation: New benchmark and state of the art

analysis. In CVPR’14. 1, 5, 6, 7, 8

[4] N. Bansal, A. Blum, and S. Chawla. Correlation clustering.

Machine Learning, 56(1–3):89–113, 2004. 2

[5] V. Belagiannis, S. Amin, M. Andriluka, B. Schiele, N. Navab,

and S. Ilic. 3D pictorial structures for multiple human pose

estimation. In CVPR’14. 2

[6] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L.

Yuille. Semantic image segmentation with deep convolutional

nets and fully connected crfs. In ICLR, 2015. 5

[7] X. Chen and A. Yuille. Articulated pose estimation by a

graphical model with image dependent pairwise relations. In

NIPS’14. 1, 2, 6, 8

[8] X. Chen and A. Yuille. Parsing occluded people by flexible

compositions. In CVPR, 2015. 2, 7, 8

[9] S. Chopra and M. Rao. The partition problem. Mathematical

Programming, 59(1–3):87–115, 1993. 2, 3

[10] E. D. Demaine, D. Emanuel, A. Fiat, and N. Immorlica. Cor-

relation clustering in general weighted graphs. Theoretical

Computer Science, 361(2–3):172–187, 2006. 2

[11] M. M. Deza and M. Laurent. Geometry of Cuts and Metrics.

Springer, 1997. 2

[12] M. Eichner and V. Ferrari. We are family: Joint pose estima-

tion of multiple persons. In ECCV’10. 2, 7

[13] G. Ghiasi, Y. Yang, D. Ramanan, and C. Fowlkes. Parsing

occluded people. In CVPR’14. 7

[14] R. Girshick. Fast r-cnn. In ICCV’15. 4, 5

[15] G. Gkioxari, B. Hariharan, R. Girshick, and J. Malik. Using

k-poselets for detecting people and localizing their keypoints.

In CVPR’14. 1, 2

[16] H. Jiang and D. R. Martin. Global pose estimation using

non-tree models. In CVPR’09. 2

[17] S. Johnson and M. Everingham. Clustered pose and nonlinear

appearance models for human pose estimation. In BMVC’10.

5, 6

[18] S. Johnson and M. Everingham. Learning Effective Human

Pose Estimation from Inaccurate Annotation. In CVPR’11. 1,

5

[19] S. Kim, C. Yoo, S. Nowozin, and P. Kohli. Image seg-

mentation using higher-order correlation clustering. TPAMI,

36:1761–1774, 2014. 2

[20] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet

classification with deep convolutional neural networks. In

NIPS’12. 5

[21] L. Ladicky, P. H. Torr, and A. Zisserman. Human pose esti-

mation using a joint pixel-wise and part-wise formulation. In

CVPR’13. 2

[22] L. Pishchulin, M. Andriluka, P. Gehler, and B. Schiele. Strong

appearance and expressive spatial models for human pose

estimation. In ICCV’13. 2, 4

[23] L. Pishchulin, A. Jain, M. Andriluka, T. Thormaehlen, and

B. Schiele. Articulated people detection and pose estimation:

Reshaping the future. In CVPR’12. 1, 2

[24] D. Ramanan. Learning to parse images of articulated objects.

In NIPS’06. 2

[25] X. Ren, A. C. Berg, and J. Malik. Recovering human body

configurations using pairwise constraints between parts. In

ICCV’05. 2

[26] B. Sapp and B. Taskar. Multimodal decomposable models for

human pose estimation. In CVPR’13. 1, 2, 5

[27] K. Simonyan and A. Zisserman. Very deep convolutional

networks for large-scale image recognition. CoRR,14. 5

[28] M. Sun and S. Savarese. Articulated part-based model for

joint object detection and pose estimation. In ICCV’11. 1, 2,

7

[29] J. Tompson, R. Goroshin, A. Jain, Y. LeCun, and C. Bregler.

Efficient object localization using convolutional networks. In

CVPR’15. 1, 2, 7

[30] J. J. Tompson, A. Jain, Y. LeCun, and C. Bregler. Joint

training of a convolutional network and a graphical model for

human pose estimation. In NIPS’14. 1, 5, 6, 7

[31] A. Toshev and C. Szegedy. Deeppose: Human pose estimation

via deep neural networks. In CVPR’14. 2, 5

[32] J. Uijlings, K. van de Sande, T. Gevers, and A. Smeulders.

Selective search for object recognition. IJCV’13. 4

[33] X. Fan, K. Zheng, Y. Lin, and S. Wang. Combining local ap-

pearance and holistic view: Dual-source deep neural networks

for human pose estimation. In CVPR’15. 6

[34] Y. Yang and D. Ramanan. Articulated human detection with

flexible mixtures of parts. PAMI’13. 2, 7

[35] J. Yarkony, A. Ihler, and C. C. Fowlkes. Fast planar corre-

lation clustering for image segmentation. In ECCV, 2012.

2

4937