Beyond Frontal Faces: Improving Person Recognition Using Multiple Cues
Ning Zhang1,2 , Manohar Paluri2 , Yaniv Taigman2 , Rob Fergus2 , Lubomir Bourdev2
1
2
UC Berkeley
Facebook AI Research

arXiv:1501.05703v2 [cs.CV] 30 Jan 2015

{nzhang}@eecs.berkeley.edu

{mano, yaniv, robfergus, lubomir}@fb.com

Abstract
We explore the task of recognizing peoples’ identities
in photo albums in an unconstrained setting. To facilitate
this, we introduce the new People In Photo Albums (PIPA)
dataset, consisting of over 60000 instances of ∼2000 individuals collected from public Flickr photo albums. With
only about half of the person images containing a frontal
face, the recognition task is very challenging due to the
large variations in pose, clothing, camera viewpoint, image
resolution and illumination. We propose the Pose Invariant
PErson Recognition (PIPER) method, which accumulates
the cues of poselet-level person recognizers trained by deep
convolutional networks to discount for the pose variations,
combined with a face recognizer and a global recognizer.
Experiments on three different settings confirm that in our
unconstrained setup PIPER significantly improves on the
performance of DeepFace, which is one of the best face recognizers as measured on the LFW dataset.

Figure 1: We are able to easily recognize people we know
in unusual poses, and even if their face is not visible. In this
paper we explore the subtle cues necessary for such robust
viewpoint-independent recognition.

nition, and second, it is not clear how to go beyond a frontal
face and leverage these subtle cues. In this paper we address
both of these problems.
We introduce the People In Photo Albums (PIPA) dataset,
a large-scale recognition dataset collected from Flickr photos with creative commons licenses. It consists of 37,107
photos containing 63,188 instances of 2,356 identities and
examples are shown in Figure 2. We tried carefully to preserve the bias of people in real photo albums by instructing annotators to mark every instance of the same identity
regardless of pose and resolution. Our dataset is challenging due to occlusion with other people, viewpoint, pose and
variations in clothes. While clothes are a good cue, they are
not always reliable, especially when the same person appears in multiple albums, or for albums where many people
wear similar clothes (sports, military events), as shown in
Figure 3. As an indication of the difficulty of our dataset,
the DeepFace system [32], which is one of the state-of-theart recognizers on LFW [18], was able to register only 52%
of the instances in our test set and, because of that, its overall accuracy on our test set is 46.66%. We plan to make the
dataset publicly available.
We propose a Pose Invariant PErson Recognition
(PIPER) method, which uses part-level person recognizers
to account for pose variations. We use poselets [2] as our
part models and train identity classifiers for each poselet.
Poselets are classifiers that detect common pose patterns. A
frontal face detector is a special case of a poselet. Other

1. Introduction
Recognizing people we know from unusual poses is easy
for us, as illustrated on Figure 1. In the absence of a clear,
high-resolution frontal face, we rely on a variety of subtle cues from other body parts, such as hair style, clothes,
glasses, pose and other context. We can easily picture Charlie Chaplin’s mustache, hat and cane or Oprah Winfrey’s
curly volume hair. Yet, examples like these are beyond
the capabilities of even the most advanced face recognizers.
While a lot of progress has been made recently in recognition from a frontal face, non-frontal views are a lot more
common in photo albums than people might suspect. For
example, in our dataset which exhibits personal photo album bias, we see that only 52% of the people have high resolution frontal faces suitable for recognition. Thus the problem of recognizing people from any viewpoint and without
the presence of a frontal face or canonical pedestrian pose is
important, and yet it has received much less attention than it
deserves. We believe this is due to two reasons: first, there is
no high quality large-scale dataset for unconstrained recog1

 (a) While clothing can be discriminative it does not help for military or busi- (b) The same individual may appear in multiple albums wearness activities, for example, where people dress similarly.
ing different clothes.

Figure 3: Challenges of our dataset. Besides significant variations in pose and viewpoint, clothing is not always a reliable
cue for person recognition in our dataset.

Facebook Co-Identification Dataset (in progress)

system.
We are inspired by the work of Zhang et al. [38], which
uses deep convolutional networks trained on poselet detected patches for attribute classification. However our
problem is significantly harder than attribute classification
since we have many more classes with significantly fewer
training examples per class. We found that combining parts
by concatenating their feature in the manner of [38] is not
effective for our task. It results in feature vectors that are
very large and overfit easily when the number of classes
is large and training examples are few. Instead, we found
training each part to do identity recognition and combining
theirwith
predictions
achieves better
performance.
Unlike [3],
•  Collected from Flickr public photos
Creative
Commons
License
we propose a new way to handle the sparsity from poselet
detections
which boosts
the performance team
by a large margin.
•  Annotated as same / not-same by
Facebook
annotations
We demonstrate the effectiveness of PIPER by using
three different experimental settings on our dataset. Our
•  Includes people / cats / dogs
method can achieve 83.05% accuracy over 581 identities on
the test set. Moreover when a frontal face is available, it improves the accuracy over DeepFace from 89.3% to 93.4%,
Figure 2: Example photos from our dataset. These are
which is ∼40%
decrease in relative error.
taken
from
a
single
album
and
show
the
associated
identiCollected
from FlickrLicense
public photos with Creative Commons
License
with• Creative
Commons
ties. Each person is annotated with a ground truth bound2. Related
ing
box around
theteam
hear,
with each color
•  Annotated
as same
/ not-same
by representing
Facebookone
annotations
teamWork
Facebook
annotations
identity. If the head is occluded, the expected position is
Face recognition There has been dramatic progress made
annotated.
•  Includes
people / cats / dogs
in face recognition in the past few decades from EigenFace

Facebook
Co-Identification Dataset (in progress)
n Dataset
(in progress)

examples are a hand next to a hip or head-and-shoulders in
a back-facing view, or legs of a person walking sideways.
A small and complementary subset of such salient patterns
is automatically selected as described in [2]. Examples of
poselets are shown in Figure 4. While each poselet is not
as powerful as a custom designed face recognizer, it leverages weak signals from specific pose pattern that is hard to
capture otherwise. By combining their predictions we accumulate the subtle discriminative information coming from
each part into a robust pose-independent person recognition

[34] to the state-of-art face recognition system [32] by using
deep convolutional nets. Most of the existing face recognition systems require constrained setting of frontal faces
and explicit 3D face alignments or facial keypoint localizations. Some other works [36, 5] have addressed robust
face recognition systems to deal with varying illumination,
occlusion and disguise. Due to our unconstrained setting,
most of conventional face recognition systems have limited
performance on our dataset.
Person identification in photo albums Tagging of personal photo albums is an active research topic. To ad-

 pose a deep network using pairs of people to encode photometric transformation. Yi et al. [37] used a siamese deep
network to learn the similarity metric between pairs of images.

Figure 4: Example of poselet activations. These are the top
4 poselet detections ranked by part weight wi mentioned in
Sec 4.3.

dress the limitation of face recognition systems, various approaches have been proposed to incorporate contexts. For
examples, the authors in [1, 25] proposed methods to incorporate contextual cues including clothing appearance and
meta data from photos for person identification in photo
collections. Sivic et al. [30] proposed a simple pictorial
structure model to retrieve all the occurrences of the same
individual in a sequence of photographs. Lin et al. [24] presented a generative probabilistic approach to model crossdomain relationships to jointly tag people, events and locations. In [13], the authors try to find all images of each
person in the scene on a set of photos from a crowed public event by integrating multiple cues including timestamps,
camera pose and co-occurrence of people.
There is also some related work to discover the social
connection between people in the photo collections. Wang
et al. [35] proposed a model to represent the relationship
between the social relationships, position and pose of people and their identities. In [12], the authors investigated the
different factors that are related to the positions of people in
a group image.
Another interesting direction is to name characters in TV
series. In [9, 29, 8], the authors proposed approach to automatically label the characters by using aligned subtitle and
script text. Tapaswi et al. [33] proposed a Markov Random Field (MRF) method to combine face recognition and
clothing features and they tried to name all the appearance
of characters in TV series including non frontal face appearance. Later they presented another semi-supervised learning method [4] for the same task.
Person re-identification in videos The task of person reidentification is to match pedestrian images from different cameras and it has important applications in video and
there are some related work for this topic. Existing work is
mainly focused on metric learning [27, 17, 22], mid-level
feature learning [16, 39, 40, 10, 20, 26]. Li et al. [23] pro-

Deep convolutional networks In the past few years, deep
convolutional networks originally pioneered by LeCun et
al. [21] have been a tremendous success by achieving the
state-of-art performance in image classification [19], object
detection [14], face recognition [32] and other computer vision tasks. The strength of the deep nets is its ability to learn
discriminative features from raw image input unlike handengineered features. DeCAF [7] showed the deep features
extracted from the network pretrained on large datasets can
be generalized to other recognition problems.

3. People In Photo Albums Dataset
To our knowledge, there is no existing large scale dataset
for the task of person recognition. The existing datasets for
person re-identification, such as VIPeR [15] and CUHK01
[22], come mostly from videos and they are low resolution
images taken from different cameras from different viewpoints. The Gallagher Collection Person Dataset [11] is the
closest to what we need, however the released subset has
only 931 instances of 32 identities which is approximately
1.5% of the size of our dataset. Furthermore, [11] have only
labeled the frontal faces. The Buffy dataset used in [29, 9]
is a video dataset and it only has less than 20 different characters.
Our problem setting is to identify person in the “wild”
without any assumption of frontal face appearance or
straight pedestrian pose. We don’t assume that the person
is detected by a detector; our instruction to annotators is to
mark the head (even if occluded) for any people they can
co-identify, regardless of their pose, and the image resolution.

3.1. General statistics
We collected our dataset, People In Photo Albums
(PIPA) Dataset, from public photo albums uploaded to
Flickr 1 and we plan to make it publicly available. All of
our photos have Creative Commons Attributions License.
Those albums were uploaded from 111 users. Photos of
the same person have been labeled with the same identity
but no other identifying information is preserved. Table 1
shows statistics of our dataset.

3.2. Collection Method
Our data collection consists of the following steps:
1. Album Filtering. After downloading thousands of albums from Flickr, we first show the annotators a set
1 https://www.flickr.com/

 Split
Photos
Albums
Instances
Identities
Avg/identity
Min/identity
Max/identity

All
37,107
1,438
63,188
2,356
26.82
5
2928

Train
17,000
579
29,223
1,409
20.74
10
99

Val
5,684
342
9,642
366
26.34
5
99

Test
7,868
357
12,886
581
22.18
10
99

Leftover
6,555
160
11,437
-

Table 1: Statistics of our dataset.

the three sets is disjoint. We do a random permutation of the uploaders and we pick the first K of them
so that the number of person instances reaches about
50% and we assign those to the training set. We assign
the next 25% to validation and the remaining 25% to
test. While we target 50-25-25 split we cannot assure
that the instances will be partitioned precisely due to
the constraints we impose. See Table 1 for more details about the splits.

4. Pose Invariant Person Recognition (PIPER)
of photos from the album and ask them to filter out
albums which are not of people albums, such as landscape, flowers, or photos where person co-occurrence
is very low.
2. Person Tagging. Then given each album, we ask the
annotators to select all the individuals that appear at
least twice in that album and draw a bounding box
around their heads with different color indicating different identity. If the head is partially/fully occluded,
we mark the region of where the head should be. The
head bounds may also be partially/fully outside the image. Not every person is tagged. In a crowd scene we
ask the annotators to tag no more than 10 people. The
interface of person tagging is shown in Fig 5a.
3. Cross-Album Merging. Often the same identities appear in multiple albums, in which case their identities should be merged. While it is not feasible to do
so across all albums, we consider the set of albums
uploaded by the same uploader and we try to find
the same identities appearing in multiple albums and
merge them. Showing all identities from all albums is
a challenging UI task for uploaders that have dozens
of large albums. We show our annotation interface in
Figure 5b.
4. Instance Frequency Normalization. After merging,
we count the number of instances for each individual and discard individuals that have less than 10 instances. In addition, a few individuals have a very large
number of instances which could bias our dataset. To
prevent such bias, we restrict the maximum number of
instances per individual to be 99. We randomly sample
99 instances and move the remaining ones into a “leftover” set. Our leftover set consists of 11,437 instances
of 54 individuals.
5. Dataset Split. We split the annotations randomly into
three sets – training, validation and test. To ensure
complete separation between the sets, all the photos
of the same uploader fall in the same set. That ensures
that the set of photos, identities and instances across

We introduce a novel view invariant approach to combine
information of different classifiers for the task of person
recognition. It consists of three components:
• The global classifier, a CNN trained on the full body
of the person.
• A set of 107 poselet classifiers, each is a CNN trained
on the specific poselet pattern using [2]. 2
• An SVM trained on the 256 dimensional features from
DeepFace [32].
In total we have 109 part classifiers. The identity prediction
of PIPER is a linear combination of the predicted probabilities of all classifiers:
X
s(X, y) =
wi Pi (y X)
(1)
i

Here Pi (y X) is the normalized probability of identity label
y given by part i for feature X and wi is the associated
weight of the part. The final identity prediction is y ∗ (X) =
argmaxy s(X, y).
Here is an overview of the training steps. The next sections
provide a more detailed description.
1. We run poselets [2] over our dataset and match the person predictions coming from poselets to our ground
truths (see Section 4.1).
2. Using the poselet patches of step 1, we train a CNN for
each poselet to recognize the identities on our training set. In addition, we train a CNN for the global
classifier using the patches corresponding to the full
body images. In all cases we use the Convolutional
Neural Net architecture by Krizhevsky et al. [19]. We
fine-tune the network pre-trained on ImageNet on the
task of identity recognition. While recent architectures
have improved the state-of-the art [28, 31] and might
further improve our performance, we decided to use
the Krizhevsky architecture because its performance
is well studied on a number of visual tasks [7]. We
2 The

original set of poselets is 150 but some of them did not train well.

 (a) Interface for annotating identities in one album.

(b) Interface for merging identities across albums.

Figure 5: Interfaces for our annotation system. The interface for annotating identities in one album in shown in (a)
where the annotator can scroll over the faces on the left. If clicking on one individual, it will show all the instances of that
person. The original images and head annotations are shown on the right where the annotators draw all the heads of different
individuals. If the person’s head is occluded, the annotator will draw a bounding box around the expected position of the
head. We show the interface for merging in (b). The top row shows a set of merged individuals. Each column in the bottom
section corresponds to an album from the same uploader. Each face is an example face of a different individual. Merging is
done by selecting a set of faces across albums, optionally selecting an individual from the top row to merge into, and clicking
the merge button.
then discard the final FC8 layer and treat the activations from the FC7 layer as a generic feature on which
we train SVMs in the next steps.
3. We partition the validation set into two halves. We
train an SVM for each part using the FC7 layer feature from Step 2 on the first half of validation and use
it to compute the identity predictions Pi (y X) on the
second half, and vice versa (see Section 4.2).
4. We use the identity predictions of all parts on the validation set to estimate the mixing components wi (see
Section 4.3).
5. We split the test set in half and train SVMs on top of
the FC7 features on the first half of the test set and use
them to compute the identity predictions Pi (y X) on
the second half, and vice versa.
6. We use the identity predictions on the test set for each
part Pi (y X) as well as the mixing components wi to
compute the combined identity prediction S(X, y) using equation 1.
In the next sections we will describe the training steps,
and way we compute Pi (y X) and wi .

4.1. Computing Part Activations
Our groundtruth annotations consist of bounding boxes of
heads. From the head boxes, we estimate the bounding box
locations by setting approximate offset and scaling factor.
We run poselets [2] on the images, which returns bounding boxes of detected people in the images, each of which
comes with a score and locations of associated poselet activations. Examples of poselet detections are shown in Figure
4. We use a bipartite graph matching algorithm to match the
ground truth bounds to the ones predicted by the poselets.
This algorithm performs globally optimal matching by preferring detections with higher score and higher overlap to
truths. The output of the algorithm is a set of poselet activations associated with each ground truth person instance.
We extract the image patches at each poselet and use them
to train part-based classifiers.

4.2. Training the Part Classifiers Pi (y X)
4.2.1

Global classifier P0 (y X)

Using the FC7 layer of the CNN trained for the full body
area of each instance, we train a multi-class SVM to predict
each identity y. We refer to its prediction as P0 (y X).

 Identities not
trained for part i

Given the FC7 layer features X extracted from detected part
i patch and identity labels y, we train a multi-class SVM on
X to predict y and we denote the softmax of the output score
as Pˆi (y X).
Notice that Pˆi is sparse in two ways:
• Each poselet activates only on instances that exhibit
the specific pose of that poselet. Some poselets may
activate on 50% while others on as few as 5% of the
data.
• Not all identities have examples for all poselets and
thus each poselet level SVM classifier for part i is only
trained on a subset Fi of all identities. Thus Pˆi (y X)
is inflated when y ∈ Fi and is zero otherwise.
The sparsity pattern is correlated with the pose of the person
and has almost no correlation to the identity that we are trying to estimate. Thus properly accounting for the sparsity is
important in order to get high accuracy identity recognition.

All instances

All identities

Replacing
with global
model
Part i
doesn’t
activate on
these
instances

All instances

Part-level SVM classifier Pˆi (y X)
All instances

4.2.2

Figure 6: Example of sparsity filling.
On the left we show
missing identities
the predictions of the global model for every identity and
every instance. The poselet classifier in the middle does
not activate for two instances (the white rows) and is not
trained to recognize two identities (the white columns). On
the right we show how in the normalized probability we
fill-in missing rows from the global model acc as in the top
of Equation 2. In addition, (not shown in the figure) we
account for the missing columns by linearly interpolating
each row with the global model based on the likelihood that
the identity is not coming from one of the missing columns.

5. Experiments
4.2.3

Sparsity filling

We address both of these sparsity issues by using the probability distribution of our global model P0 which is defined
for all identities and activates on all instances:
(
P0 (y X), if part i doesn’t activate, or
Pi (y X) =
P (y ∈ Fi )Pˆi (y X) + P (y ∈
/ Fi )P0 (y X)
(2)
P (y ∈ Fi ) =

X

0

P0 (y  X)

(3)

y 0 ∈Fi

In Figure 6 we give a visual intuition behind this formula.

4.3. Computing the part weights wi
We use the validation set to compute w. We split the validation set into two equal subsets. We train the part-based
SVMs on one subset and use them to compute Pi (y X)
on all instances of the second subset, and vice versa. Let
Pij (y X) denote the probability that the classifier for part
i assigns to instance j being of class y given feature X.
We formulate a binary classification problem which has
one training example per pair of instance j and label y. If
we have K parts its feature vector is K + 1 dimensional:
[P0j (y X); P1j (y X); ... Pkj (y X)]. Its label is 1 if instance
j’s label is y and -1 otherwise. We solve this by training a
linear SVM. The weights w are the weights of the trained
SVM.
We use the first split of validation to do a grid search for
the C parameter of the SVM and test on the second half.
Once we find the optimal C we retrain on the entire validation set to obtain the final vector w.

We report results of our proposed method on our PIPA
dataset and compare it with baselines. Specifically, we
conduct experiments in three different settings: 1) Person
recognition, 2) One-shot person identification, and 3) Unsupervised identity retrieval.
In all experiments we use the training split of our dataset
to train the deep networks for our global model and each
poselet and we use the validation set to compute the mixing
weights w and tune the hyper-parameters. All of our results
are evaluated on the test split.

5.1. Person Recognition
We first present experimental results on the person
recognition task with our PIPA dataset. It is a standard supervised classification task as we train and test on same set
of identities. Since the set of identities between training and
test sets is disjoint, we split our test set in two equal subsets.
We train an SVM on the first, use it to compute Pi (y X)
on the second and vice versa. We then use the weights w
trained on the validation set to get our final prediction as
the identity that maximizes the score in equation 1 and we
average the accuracy from both halves of the test set. Qualitative examples are shown on Figure 7.
5.1.1

Overall Performance

Table 2 shows the accuracy in this setting compared to some
standard baselines. We compared it against DeepFace [32],
which is one of the state-of-the-art face recognizers. Although it is very accurate, it is a frontal face recognizer, so
it doesn’t trigger on 48% of our test set and we use chance

 cropped'bounding'box'

cropped'bounding'box'

(a) Body falls outside image

(b) Unusual pose

(c) Same clothing similar pose

(d) Confusion with other person

Figure 7: Examples that the combination of the Global model and DeepFace misclassify and are recovered by using all of
PIPER. (a) In a closeup shot the full body falls outside the image and the extracted full-body patch, shown on the right, is
severely misaligned. A profile-face poselet should handle this case without misalignment. (b) In unusual pose the full body
patch may fall on the background or (d) on another person which will further confuse the classifier. In (c) people have the
same clothes and similar pose which will confuse the global model.
performance for those instances. As a second baseline we
trained an SVM using the FC7 features of the CNN proposed by Krizhevsky et al. and pretrained on ImageNet[6].
We use its activations after showing it the full body image
patch for each instance. The Global Model baseline is the
same CNN, except it was fine-tuned for the task of identity
recognition on our training set. We also tested the performance of our model by combining the sparse part predictions, i.e. using Pˆi (y X) instead of Pi (y X) in equation 1.
The performance gap of more than 7% shows that sparsity
filling is essential to achieve high recognition accuracy.
5.1.2

Ablation Study

Our method consists of three components – the fine-tuned
Krizhevsky CNN (the Global Model), the DeepFace recognizer, and a collection of 108 Poselet-based recognizers. In

Method
Chance Performance
DeepFace [32]
FC7 of Krizhevsky et al. [19]
Global Model
PIPER w/out sparsity filling
PIPER

Classification accuracy
0.17%
46.66%
56.07%
67.60%
75.35%
83.05%

Table 2: Person recognition results on PIPA test set using
6442 training examples over 581 identities

this section we explore using all combinations of these three
components3 . For each combination we retrain the mixture
3 Since our method relies on sparsity filling from a global model
P0 (y X), to remove the effect of the global model we simply set P0 (y X)
to be uniform distribution.

 Global Model
X
–
–
X
X
–
X

DeepFace[32]
–
X
–
X
–
X
X

Poselets
–
–
X
–
X
X
X

Accuracy
67.60%
46.66%
72.18%
79.95%
78.79%
78.08%
83.05%

Table 3: Person recognition performance on the PIPA test
set using 6442 training examples over 581 identities as we
disable some of the components of our method. PIPER gets
more than 3% gain over the very strong baseline of using
the fine-tuned CNN combined with the DeepFace model.
DeepFace’ s score is low because it only fires on 52% of the
test images and we use chance performance for the rest.
Method
Global Model
DeepFace[32]
PIPER

Non-faces subset
64.3%
0.17%
71.8%

Figure 8: Recognition accuracy as a function of number of
training examples per identity, with σ = 1 error bar. As
we increase the number of training examples our system’s
accuracy grows faster than the full-body baseline. Chance
performance is 0.0017.

Faces subset
70.6%
89.3%
93.4%

Table 4: Performance on the test set when split into the subset where frontal face is visible (52%) and when it is not
(48%).
weights w and re-tune the hyper parameters. Table 3 shows
the performance of each combination of these components.
As the table shows, the three parts PIPER are complementary and combining them is necessary to achieve the best
performance.
5.1.3

Performance on face and non-face instances

Since the presence of a high resolution frontal face provides a strong cue for identity recognition and allows us
to use the face recognizer, it is important to consider the
performance when a frontal face is present and when it is
not. Table 4 shows the performance on the face and nonface part of our test set. We considered the instances for
which DeepFace generated a signature as the face subset.
As the figure shows, when the face is not present we can significantly outperform a fine-tuned CNN on the full image.
More importantly, the contextual cues and combinations of
many classifiers allow us to significantly boost the recognition performance even when a frontal face is present.

5.2. One-Shot Learning
Figure 8 shows the performance of our system when the
training set is tiny. We randomly pick one, two or three instances of each identity in our test set, train on those and
report results on the rest of the test set. Our system performs very well even with a single training example per

Figure 9: Performance of our method on identity retrieval.

identity, achieving 28.1% accuracy for our test set of 581
identities. This result illustrates the powerful generalization
capability of our method. The generalization capabilities of
deep features are well studied, but we believe here we are
also helped by the mixture of multiple part-based classifiers,
since our system improves faster than the global method of
fine-tuned Krizhevsky’s CNN.

5.3. Unsupervised identity retrieval
We evaluate our method on the task of retrieval: Given
an instance, we measure the likelihood that one of the K
nearest neighbors will have the same identity.
To do this we used the SVMs trained on split 0 of the
validation set to predict the 366 identities in the validation
set. We applied them to the instances in the test set to obtain
a 366-dimensional feature vector for each part and we combine the part predictions using equation 1 with w trained on
the validation set to obtain a single 366-dimensional feature
for each instance in the test set. We then, for each instance
of the test set, compute the K nearest neighbors using Eucledian distance and we consider retrieval as successful if at
least one of them is of the same identity. This has the effect

 Query

Top 1

Top 2

Top 3

Top 4

Top 5

ample and performs surprisingly well at the task of image
retrieval. While we have used PIPER for person recognition, the algorithm readily applies to generic instance coidentification, such as finding instances of the same car or
the same dog. We introduced the People In Photo Albums
dataset, the first of its kind large scale data set for person
coidentification in photo albums. We hope our dataset will
steer the vision community towards the very important and
largely unsolved problem of person recognition in the wild.
Acknowledgments We would like to thank Nikhil Johri
for collecting dataset and developing the annotation interfaces. We also would like to thank Trevor Darrell for useful
discussions on the ideas of the paper.
Copyrights The
copyright
references
of
all
the images shown in the paper are in http:
//www.eecs.berkeley.edu/˜nzhang/piper_
image_license.pdf.

References

Figure 10: Example of PIPER results on unsupervised identity retrieval. For each row, the left one is the query image,
then the top 5 ranked retrieval images. Those are cropped
bounding box images on test split.

of using the identities in the validation set as exemplars and
projecting new instances based on their similarities to those
identities. As Figure 9 shows our method is quite effective
on this task – despite the low dimensional feature and without any metric learning, the nearest neighbor of 64% of the
examples is of the same identity as the query. If we use the
predictions of the Krizhevsky’s CNN trained on ImageNet
and fine-tuned on our training set, which is known to be a
very powerful baseline, the nearest neighbor is of the same
class in only 50% of the examples. This suggests that our
model is capable of building a powerful and compact identity vector independent of pose and viewpoint. Examples of
our method are shown in Figure 10.

6. Conclusion
We described PIPER, our method for viewpoint and
pose independent person recognition. We showed that
PIPER significantly outperforms our very strong baseline
– combining a state-of-the-art CNN system fine-tuned on
our dataset with a state-of-the-art frontal face recognizer.
PIPER can learn effectively even with a single training ex-

[1] D. Anguelov, K. chih Lee, S. B. Gktrk, and B. Sumengen.
Contextual identity recognition in personal photo albums. In
CVPR, 2007. 3
[2] L. Bourdev and J. Malik. Poselets: Body part detectors
trained using 3D human pose annotations. In International
Conference on Computer Vision (ICCV), 2009. 1, 2, 4, 5
[3] L. D. Bourdev, S. Maji, and J. Malik. Describing people: A
poselet-based approach to attribute classification. In ICCV,
2011. 2
[4] M. Buml, M. Tapaswi, and R. Stiefelhagen. Semi-supervised
learning with constraints for person identification in multimedia data. In CVPR’13, 2013. 3
[5] Z. Cui, W. Li, D. Xu, S. Shan, and X. Chen. Fusing robust
face region descriptors via multiple metric learning for face
recognition in the wild. In CVPR, 2013. 2
[6] J. Deng, W. Dong, R. Socher, L. jia Li, K. Li, and L. Fei-fei.
Imagenet: A large-scale hierarchical image database. In In
CVPR, 2009. 7
[7] J. Donahue, Y. Jia, O. Vinyals, J. Hoffman, N. Zhang,
E. Tzeng, and T. Darrell. Decaf: A deep convolutional activation feature for generic visual recognition, 2014. 3, 4
[8] M. Everingham, J. Sivic, and A. Zisserman. “Hello! My
name is... Buffy” – automatic naming of characters in TV
video. In Proceedings of the British Machine Vision Conference, 2006. 3
[9] M. Everingham, J. Sivic, and A. Zisserman. Taking the bite
out of automated naming of characters in tv video. Image
Vision Comput., 27(5):545–559, Apr. 2009. 3
[10] M. Farenzena, L. Bazzani, A. Perina, V. Murino, and
M. Cristani. Person re-identification by symmetry-driven accumulation of local features. In CVPR, 2010. 3
[11] A. Gallagher and T. Chen. Clothing cosegmentation for recognizing people. In Proc. CVPR, 2008. 3

 [12] A. C. Gallagher and T. Chen. Understanding images of
groups of people. In CVPR, 2009. 3
[13] R. Garg, D. Ramanan, S. Seitz, and N. Snavely. Where’s
waldo: Matching people in images of crowds. In IEEE
Conference on Computer Vision and Pattern Recognition
(CVPR), 2011. 3
[14] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic
segmentation. arXiv preprint:1311.2524. 3
[15] D. Gray, S. Brennan, and H. Tao. Evaluating appearance
models for recognition, reacquisition, and tracking. In In
IEEE International Workshop on Performance Evaluation
for Tracking and Surveillance, Rio de Janeiro, 2007. 3
[16] D. Gray and H. Tao. Viewpoint invariant pedestrian recognition with an ensemble of localized features. In Proceedings
of the 10th European Conference on Computer Vision: Part
I, ECCV ’08, pages 262–275, 2008. 3
[17] M. Guillaumin, J. Verbeek, and C. Schmid. Is that you? metric learning approaches for face identification. In ICCV, Kyoto, Japan, Sept. 2009. 3
[18] G. B. Huang, M. Ramesh, T. Berg, and E. Learned-Miller.
Labeled faces in the wild: A database for studying face
recognition in unconstrained environments. Technical Report 07-49, University of Massachusetts, Amherst, October
2007. 1
[19] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet
classification with deep convolutional neural networks. In
Neural Information Processing Systems (NIPS), 2012. 3, 4,
7
[20] R. Layne, T. Hospedales, and S. Gong.
Person reidentification by attributes. In BMVC”12, 2012. 3
[21] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E.
Howard, W. Hubbard, and L. D. Jackel. Backpropagation
applied to handwritten zip code recognition. Neural Comput., 1(4):541–551, Dec. 1989. 3
[22] W. Li, R. Zhao, and X. Wang. Human reidentification
with transferred metric learning. In ACCV, volume 7724 of
Lecture Notes in Computer Science, pages 31–44. Springer,
2012. 3
[23] W. Li, R. Zhao, T. Xiao, and X. Wang. Deepreid: Deep
filter pairing neural network for person re-identification. In
CVPR’14, 2014. 3
[24] D. Lin, A. Kapoor, G. Hua, and S. Baker. Joint people, event,
and location recognition in personal photo collections using
cross-domain context. In Proceedings of the 11th European
Conference on Computer Vision: Part I, ECCV’10, pages
243–256, 2010. 3

[25] M. Naaman, H. Garcia-Monlina, A. Paepcke, and R. B. Yeh.
Leveraging context to resolve identity in photo albums. In
JCDL’05, 2005. 3
[26] O. Oreifej, R. Mehran, and M. Shah. Human identity recognition in aerial images. In CVPR, pages 709–716, 2010. 3
[27] B. Prosser, W. shi Zheng, S. Gong, and T. Xiang. Person reidentification by support vector ranking. In BMVC’10, 2010.
3
[28] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition, 2014. 4
[29] J. Sivic, M. Everingham, and A. Zisserman. “Who are you?”
– learning person specific classifiers from video. In Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition, 2009. 3
[30] J. Sivic, C. L. Zitnick, and R. Szeliski. Finding people in
repeated shots of the same scene. In BMVC’06, 2006. 3
[31] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed,
D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich.
Going deeper with convolutions, 2014. 4
[32] Y. Taigman, M. Yang, M. Ranzato, and L. Wolf. DeepFace:
Closing the Gap to Human-Level Performance in Face Verification. In Conference on Computer Vision and Pattern
Recognition (CVPR), 2014. 1, 2, 3, 4, 6, 7, 8
[33] M. Tapaswi, M. Buml, and R. Stiefelhagen. Knock! knock!
who is it? probabilistic person identification in tv-series. In
CVPR’12, 2012. 3
[34] M. Turk and A. Pentland. Eigenfaces for recognition. J.
Cognitive Neuroscience, 3(1):71–86, Jan. 1991. 2
[35] G. Wang, A. Gallagher, J. Luo, and D. Forsyth. Seeing people in social context: Recognizing people and social relationships. In Proceedings of the 11th European Conference on
Computer Vision: Part V, ECCV’10, pages 169–182, 2010.
3
[36] J. Wright, A. Y. Yang, A. Ganesh, S. S. Sastry, and Y. Ma.
Robust face recognition via sparse representation. IEEE
Trans. Pattern Anal. Mach. Intell., 31(2):210–227, Feb.
2009. 2
[37] D. Yi, Z. Lei, and S. Z. Li. Deep metric learning for practical
person re-identification, 2014. 3
[38] N. Zhang, M. Paluri, M. Ranzato, T. Darrell, and L. D. Bourdev. PANDA: pose aligned networks for deep attribute modeling. CVPR, 2014. 2
[39] R. Zhao, W. Ouyang, and X. Wang. Unsupervised salience
learning for person re-identification. In CVPR’13, 2013. 3
[40] R. Zhao, W. Ouyang, and X. Wang. Learning mid-level filters for person re-identification. In CVPR’14, 2014. 3