Classiﬁcation-driven Active Contour for Dress Segmentation

Lixuan Yang

1,2

, Helena Rodriguez

, Michel Crucianu

and Marin Ferecatu

Shopedia SAS, 55 rue La Bo

etie, 75008, Paris, France

CEDRIC Lab, Conservatoire National des Arts et M

etiers, 292 Rue Saint-Martin, 75003, Paris, France

Keywords:

Dress Extraction, Clothing Extraction, Segmentation, SVM Classiﬁcation, Active Contour.

Abstract:

In this work we propose a a dedicated object extractor for dress segmentation in fashion images by combining

local information with a prior learning. First, a person detector is applied to localize sites in the image that

are likely to contain the object. Then, an intra-image two-stage learning process is developed to roughly

separate foreground pixels from the background. Finally, the object is ﬁnely segmented by employing an

active contour algorithm that takes into account the previous segmentation and injects speciﬁc knowledge

about local curvature in the energy function. The method is validated on a database of manually segmented

images. We show examples of both successful segmentation and difﬁcult cases. We quantitatively analyze

each component and compare with the well-known GrabCut foreground extraction method.

1 INTRODUCTION

Although the interest in developing dedicated search

engines for fashion databases is several decades old

(King and Lau, 1996), the ﬁeld only started to de-

velop with the recent massive proliferation of fash-

ion web-stores and online retail shops. Indeed, more

and more users expect online advertising to propose

items that truly correspond to their expectations in

terms of design, manufacturing and suitability. Lo-

calizing, extracting and tracking fashion items during

web browsing allows the professionals to better un-

derstand the users’ preferences and design web inter-

faces that make for a better web shopping experience.

In this work we propose a method to segment

dresses in fashion images, which is the ﬁrst step

to solving a more difﬁcult problem inspired by the

need of professionals of online advertising and fash-

ion media: present to the users relevant items from a

database of clothes, based on the content of the web

application they are consulting and its context of use.

This goes far beyond the needs of a search engine: the

user is not asked to interact with any search interface

or formulate a query, but instead she is accompanied

by automatic suggestions presenting in a non intrusive

way a selection of products that are likely to interest

her.

Many recent research efforts regarding fashion

images deal with a quite different use case, that of

meta search engines federating and comparing sev-

eral online shops. These efforts focus on improv-

ing existing search engines to help users ﬁnd prod-

ucts that match their preferences while preserving the

“browsing” aspect (Datta et al., 2008). Online shops

sometimes provide image tags for common visual at-

tributes, such as color or pattern, but they usually

form a proprietary, non heterogeneous and non stan-

dardized vocabulary, usually too small to characterize

the visual diversity of desired clothing (Redi, 2013;

Di et al., 2013). Moreover, in many cases users look

for characteristics expressed by very subjective con-

cepts and words, to describe a style, a given brand

or a speciﬁc design. For this reason, much recent re-

search work is focused in the development of detec-

tion, recognition and search of fashion items based in

visual characteristics (Datta et al., 2008; Lew et al.,

2006; Liu et al., 2007; Liu et al., 2012a).

Another approach models the target item based on

attribute selection and high-level classiﬁcation (De-

selaers et al., 2008). In (Di et al., 2013), the au-

thors train attribute classiﬁers on ﬁne-grained cloth-

ing styles, formulating image retrieval as a classiﬁ-

cation problem, by ranking items that contains the

same visual attributes as the input, which can be a

list of words or an image. A similar idea is explored

in (Hsu et al., 2011) where a set of features such as

color, texture, SIFT features and object outlines are

used to determine similarity scores between pairs of

images. In (Chen et al., 2012), the authors propose

to extract low-level features in a pose-adaptive man-

Yang, L., Rodriguez, H., Crucianu, M. and Ferecatu, M.

Classiﬁcation-driven Active Contour for Dress Segmentation.

DOI: 10.5220/0005721000220029

In Proceedings of the 11th Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2016) - Volume 4: VISAPP, pages 22-29

ISBN: 978-989-758-175-5

ner and combine complementary features for learning

attribute classiﬁers by exploring mutual dependencies

between the attributes by conditional random ﬁelds.

To narrow the semantic gap between the low-level

features of clothing and the high-level categories, (Liu

et al., 2012a) propose to adopt middle-level clothing

attributes (e.g., clothing category, color, pattern) as a

bridge. More speciﬁcally, the clothing attributes are

treated as latent variables in a latent Support Vector

Machine (SVM) recommendation model. To perform

on larger ﬁne-grained clothing attributes, (Chen et al.,

2015) proposed a novel double-path deep domain

adaptation network for attribute prediction by mod-

eling the data jointly from the unconstrained photos

and images from large-scale online shopping stores.

A second approach consists in using part-based

models to compensate the lack of pose estimation

and model complex interactions in deformable ob-

jects (Felzenszwalb et al., 2010). To predict human

occupations, (Song et al., 2011) use part-based mod-

els to characterize complex details and variable ap-

pearances of human clothing on the automatically

aligned patches of human body parts, described by

sparse coding and noise-tolerant capacities. A similar

part-based model is proposed in (Nguyen et al., 2012)

where image patches are described by a mixture of

color and texture features. Parts are also used in (Liu

et al., 2012b) to reduce the “feature gap” caused by

human pose discrepancy, by using graph parsing tech-

nique to align human parts for cloth retrieval.

Another approach is based on segmentation and

aggregation to select different cloth categories. In

(Kalantidis et al., 2013), articulated pose estimation

is followed by an over-segmentation of the relevant

parts. Then, clustering by appearance creates a refer-

ence frame that facilitates rapid classiﬁcation without

requiring an actual training step. Body joints are in-

corporated in (Jammalamadaka et al., 2013) by esti-

mating their prior distributions and then learning the

cloth–joint co-occurrences of different cloth types as

a part of a conditional random ﬁeld framework to seg-

ment the image into different clothes. A similar idea

is proposed in (Yamaguchi et al., 2015b), they formu-

late a CRF by inter-object or inter-attribute compati-

bility. (Simo-Serra et al., 2014) has exploited another

way to formulate a CRF by taking into account the

garment’s priors, 2D body pose condition and limb

segment. The framework in (Yamaguchi et al., 2015a)

is based on bottom-up clothing parsing from seman-

tic labels attached to each pixel. Local models of

clothing items are learned on-the-ﬂy from retrieved

examples and parse mask predictions are transferred

from these examples to the query image. Face de-

tection is used in (Yang and Yu, 2011) to locate and

track human faces in surveillance videos, then cloth-

ing is extracted by Voronoi partitioning to select seeds

for region growing. For the video applications, (Liu

et al., 2014) use Sift Flow and super-pixel matching to

build correspondences across frames and exploit the

cross-frame contexts to enhance human pose estima-

tion. Also for the human parsing and pose estimation,

(Dong et al., 2014) proposed an uniﬁed framework to

formulate the problem jointly via a tailored And-Or

graph.

(a) (b) (c)

Figure 1: We aim to produce a precise segmentation of the

fashion items as in (b). Recent state of the art (Yamaguchi

et al., 2015a) produces the result in right ﬁgure (c), which

is insufﬁcient to provide a precise description of the object

(dress in this case).

Unlike the methods described above, our proposal

aims to segment precisely the object of interest from

the background (foreground separation, see Fig. 1(b)),

a difﬁcult problem without user interaction and with-

out using an extensive training database. Indeed, to

propose meaningful results in terms of high-level ex-

pectations (such as product style or design) we need

to achieve a good description of the visual appear-

ance, which is much better if the object is segmented

to eliminate the effect of mixing with the background

and to include the shape outline in the description.

As an example, we show in Fig. 1(c), the results

obtained on the left image by using state-of-the art

method from (Yamaguchi et al., 2015a). Even though

this method is capable of multi-label segmentation,

the result of the segmentation is inadequate for a ﬁne

description of the dress. Instead, we aim to obtain

a ﬁner description like the one in Fig. 1(b) (which,

incidentally, is produced by the method described in

this paper).

To achieve our goal, we combine a person detec-

tor with a two stage SVM classiﬁcation to achieve a

rough estimation of the cloth contour (separation of

the object from the background). The result is then

used to seed an active contour, ﬁne-tuned by prior spe-

ciﬁc knowledge of the object structure.

The novelty of our method is twofold: ﬁrst, we

Classiﬁcation-driven Active Contour for Dress Segmentation

combine local information in the image with a learn-

ing prior to guide the segmentation (which allows to

predict the contour in the presence of small occlu-

sions, rather than just following local contours) and

second, we inject speciﬁc knowledge about the object

(local curvature) in the energy function that guides

the convergence of the active contour, which helps

disambiguate the object’s contour in cluttered back-

ground. The procedure requires a training database

for the person detector and for the foreground detec-

tion stage, together with the collection of prior knowl-

edge regarding the object to be segmented. To keep

the presentation uncluttered, in this paper we focus on

dresses, a challenging class to segment because of the

complexity of the object and the variety of the envi-

ronment in real images. However, our procedure can

be adapted to any fashion cloth item by following a

similar development approach.

Evaluation tests performed on a database of 200

manually segmented images show very promising re-

sults. We provide examples of successful segmenta-

tion, analyze difﬁcult cases, and also evaluate quanti-

tatively each component. Because existing methods,

to our knowledge, do not segment precisely fashion

items at this moment, we compare our method to the

GrabCut (Rother et al., 2004) foreground extraction

method, which is well-known in our community, is

frequently used as a baseline case in many research

works and has an open source implementation

The rest of the paper is organized as follows: In

Sec. 2.1 we give an overview of our proposal, fol-

lowed by a detailed presentation of each component:

person detector in Sec. 2.2, SVM-based detection in

Sec. 2.3 and active contour in Sec. 2.4. After the ex-

perimental validation is presented in Sec. 3, we con-

clude the paper in Sec. 4 by a discussion of the main

points and perspectives for further extensions.

2 OUR PROPOSAL

Detecting dresses in images is a difﬁcult problem be-

cause the object is deformable, can be composed of a

large variety of materials, textures and patterns, and

shows great differences in style and design inside this

class. Also, it can appear against very different and

complex backgrounds.

2.1 Overview of the Approach

Since we aim to ﬁnd as precisely as possible the con-

tour of the dress, a direct approach is deemed to fail

http://opencv.org/

because of the aforementioned difﬁculties. Instead,

we adopt a three stage method, each step preparing

the following one as follows:

1. Person Detector. We ﬁrst train a person detector

on a manually annotated database to ﬁnd the regions

of the image most likely to contain the contour of the

object. We use the articulated human detection model

with a ﬂexible mixture of parts presented in (Yang and

Ramanan, 2013), which works well for both person

detection and pose estimation, and has been tested

with success in several other fashion-related works

(see Sec. 1). The output of the person detector is a

set of parts (rectangular boxes) centered on the body

joints and oriented correctly.

2. Coarse Foreground Detection. We employ the

training data to estimate a probability map that each

pixel inside a box belongs to the object. The map

is used to seed a one-class SVM estimating the sup-

port of the distribution of positive examples (pixels

that belong to the object). Then, a two-class SVM

is trained to improve the (coarse) detection of pixels

belonging to the object, taking as negative examples

random background pixels.

3. Active Contour. The result of the two-class SVM

is used as input to a two step active contour procedure

that produces the ﬁnal segmentation. We include sev-

eral speciﬁc terms in the energy function that guides

the active contour: the ﬁrst term uses the results of the

learning stage to push the contour towards the SVM

separation frontier (i.e., the contour of the object ac-

cording to the SVM), the second term takes into ac-

count the local curvature weighted by the location on

the object. This ensures a good balance between the

local pixel behavior and the information injected by

learning, producing good results in most situations.

2.2 Person Detector

For clothing detection and recognition, applying a

person detector is a reasonable starting point. As in

many other studies (e.g. (Liu et al., 2012b; Kalan-

tidis et al., 2013; Yamaguchi et al., 2015a)), we use

the person detection with articulated pose estimation

algorithm from (Yang and Ramanan, 2013), which

has been extensively tested and proved to be very ro-

bust in many practical situations. It is based on a de-

formable model that deﬁnes the object as a combi-

nation of parts (Yang and Ramanan, 2013). The de-

tection score is deﬁned as the ﬁt to parts minus the

parts deformation cost. The mixture model not only

encodes object structure but also captures spatial re-

lations between part locations and co-occurrence re-

lations between parts.

To train the person detector, we manually annotate

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

(a) (b) (c) (d) (e)

Figure 2: Different stages of our approach: (a) output of the person detector, (b) output of the one class SVM, (c) and (d) input

and output of the two class SVM, (e) ﬁnal result after ﬁrst (green) and second (red) active contour steps.

a database of 100 images of dresses. Each person in

a training image is annotated with 14 joint points by

marking the articulations and main body parts. The

legs are usually covered by the long dresses, so the

lower parts are placed on the edges of the dress rather

than on the legs. This not only improves detection ac-

curacy, but also hints to the location of the contours.

Given an image I, the person detector provides a set

of 26 body parts, each part being a square region re-

sized to 40 × 40 pixels. Each part has a symmetric

part with respect to the vertical axis and corresponds

to a body part or articulation. In Fig. 2(a) we see

the output of the person detector on an unannotated

image. Note how boxes slightly overlap at each end

most of the time. To reduce the search space, we take

advantage of the fact that the dress contour is located

inside the boxes and we close the outline by adding

the green (upper) and black (bottom) boxes to make

the connection between the left and right body part

estimations. These new boxes are computed to ﬁt the

internal and external hull of the existing ones. All

subsequent processing of each image is done only in

the region outlined by the boxes. Probability Map.

For each location (pixel) in every box, we compute

the probability of occurrence of the dress. This prob-

ability map indicates for each box where the dress is

more likely to be found. We use this map to harvest

positive examples for the next stage (SVM classiﬁca-

tion, see Sec. 2.3). To compute the map, we manually

segment the dress in all the training images and then,

for each pixel position in each box b, we compute the

value:

(i, j) =

∑

k=1

δ(I

(i, j) ∈ Dress)

where n is the number of training images and

δ(I

(i, j) ∈ Dress) is 1 iff pixel (i, j) from box b in

image I

belongs to the dress.

2.3 Coarse Foreground Detection

In the second stage, for each box we train a two-class

SVM (Scholkopf and Smola, 2001) to separate fore-

ground pixels (dress) from the background by using

the prior information given by the probability map

and the person detector. Each pixel is described by the

RGB coordinates concatenated with other local char-

acteristics as described in Sec. 3. In a new unseen im-

age we only have the result of the person detector to

start with (the 26 part boxes). Using directly the prob-

ability map computed earlier to ﬁnd positive training

examples by thresholding does not yield enough good

examples to guarantee correct generalization. Instead,

we use the 100 most probable pixels in the box (ac-

cording to the probability map) to train a one-class

SVM that computes the support of the positive exam-

ples in the input space. We observed experimentally

that a number larger than 100 reduces the generaliza-

tion ability of the resulting classiﬁer. The main source

of problems here is that pixel classiﬁcation is prone to

local instability in cluttered scenes. To counter this,

we use the context of each part to inject conﬁdence

into the decision by allowing the neighboring parts to

vote. Concretely, for each part, we build four one-

class SVMs: one for the part itself, one for the sym-

metric part w.r.t. the vertical and two for the lower

and upper neighboring parts (see Fig. 2(b)). A pixel

is considered positive is all four one-class SVMs val-

idate it.

Merely using the output of the previous one-class

classiﬁcation fails to isolate properly the dress be-

cause pixels from the background and from the dress

may have similar descriptors. We thus take some

of the background pixels as negative examples for a

two class SVM. More precisely, the pixels predicted

by the one-class classiﬁer as belonging to the dress

are considered as positive examples. We randomly

take as negative examples an equal number of pix-

Classiﬁcation-driven Active Contour for Dress Segmentation

els from the background (outside the envelope of all

the parts) to obtain a balanced training problem. We

also include the head part in the negative examples.

In Fig. 2(c) we illustrate the training set for the two-

class SVM and in Fig. 2(d) we show the result of pre-

diction. It can be seen that the cloud of positive pre-

dictions outlines the dress quite closely, meaning the

learning problem is well posed.

2.4 Active Contour

The score of the two-class SVM on a pixel indicates

the likelihood of the dress presence. To get the ﬁnal

dress segmentation, we use the active contour (AC)

introduced in (Chan and Vese, 2001), a model that

can segment objects whose boundaries are not nec-

essarily well-supported by gradient information. The

AC minimizes an energy that drives the evolution of

the active contour towards the desired boundary. The

energy of the contour is deﬁned by its length, area and

a ﬁtting term:

F(c

,C) = µ ∗Length(C) + ν ∗ Area(inside(C))+

(1)

+λ

inside(C)

(x,y) − c

+λ

outside(C)

(x,y) − c

where C is the current contour, c

and c

are the aver-

age pixel gray level values u(x, y) inside and respec-

tively outside the contour C. The curvature term is

controlled by µ and the ﬁtting terms by λ

and λ

To achieve a faster convergence to the ﬁnal result,

a two-step procedure is employed. A ﬁrst AC is ini-

tialized with the external envelope of the parts pro-

duced by the person detector and converges rapidly

to an approximation of the desired boundary. It takes

as input the binary image produced by the two-class

SVM and requires a small µ allowing for strong cur-

vatures. Then, a second AC is used to converge to

the real contour. It is initialized with the contour pro-

duced by the previous step and has as input the gray

level image. This step is using prior information about

the object: the curvature in various areas is expected

to be different. In some images the hand will cover the

dress, we thus give a large value to µ in those parts to

obtain a smoothed result. On the contrary, the shoul-

der, the lower part of the dress and the elbow may

have strong curvatures, so we set a small value to µ

in these parts to correctly follow the contour. For the

other parts, we take a medium value for µ.

The mean values c

and c

are computed usually

on the entire image. Because of the large variability

of the background in real images, this values is mean-

ingless in a local context in our case and we replace

them by the average values calculated in a local win-

dow of size 40 × 40 pixels around each contour pixel.

To reinforce the role of the SVM-based classiﬁca-

tion on the position of the AC, we include a new term

in the energy function (Eq. 1) that pushes the contour

towards the SVM separation frontier:

svm

on(C)

| f

svm

(x,y)|

(2)

where f

svm

is the two-class SVM decision function

mapped between 0 and 1 by the logistic function

1/(1 + e

−|·|

3 EXPERIMENTAL RESULTS

At this time, we found no databases dedicated to

evaluating cloth extraction/segmentation from natural

images that could be used with our framework, i.e.

which has enough number of dresses (our object of

interest) to make training feasible. The closest we

found is Fashionista (Yamaguchi, 2015; Yamaguchi

et al., 2013) which is build to test accuracy of multi-

label assignment to pixels, but their method (Paper

Doll Parsing) trained on this database did not perform

very well for extracting long dresses (see Fig. 1 and

Fig. 3 for some examples).

(a) (b) (c)

Figure 3: Paper Doll (Yamaguchi et al., 2013) is inadequate

for the precise segmentation needed by our use case: (a)

original image, (b) our method, (c) Paper Doll.

Instead, we evaluate our method on a database of

200 manually segmented dress images, half of which

are used for training and the other half for testing.

This is enough for preliminary testing of the method,

but of course a larger database is needed for full vali-

dation, also including other fashion objects. We plan

to do this in an extension of the present work.

For the SVM prediction stage, we describe each

pixel by its RGB values concatenated with the x and

y derivatives. To correctly separate the dress from

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

(a) (b) (c) (d) (e) (f)

Figure 4: Qualitative evaluation: (a, c, e) original images; (b, d, f) segmentation results.

(a) Original image

(b) Our method (c) GrabCut

Figure 5: Visual comparison with GrabCut.

the background, especially when the two have simi-

lar colors, we further concatenate with the frequency

distribution of the power spectral density computed

in a patch of size 8 × 8 pixels around the pixel. This

is a well-known texture descriptor (Deselaers et al.,

2008). Concerning the SVM, we used the LibSVM

implementation (Chang and Lin, 2011) with C = 100

and the Gaussian kernel with scale γ = 0.1, parame-

ters obtained by cross-validation on the training data.

In regard to the computational efﬁciency of our

method, the time needed to extract the object contour

from an image of size 800x600 pixels in of the order

of 5 seconds on an average PC. This could likely be

improved by a factor of 5 to 10 by parallel computa-

tion and code optimization. However, in our applica-

tion scenario this is not needed, because the extraction

is not real-time, so we did not pursue further this di-

rection.

Quantitative Evaluation: We compare the seg-

mentation produced by our method to the one pro-

vided by a human. As performance measures we use

the average rates of true positive (TP), true negative

(TN), false positive (FP) and false negative (FN) pix-

els, traditionally used for classiﬁers. To show the im-

pact of each component, in Table 1 we compare the

segmentation obtained by the original AC (Chan and

Vese, 2001), the classiﬁcation obtained by the SVM

with and without the original AC and our method

(SVM classiﬁcation followed by the two step active

contour with energy function enhanced by curvature

compensation and SVM regularization terms). We see

that both the active contour alone and the SVM alone

produce few false positives, but a very high rate of

FN, i.e. tend to miss-classify dress regions as back-

ground. The full method achieves a more adequate

behavior in most cases: rather low error rates because

of the curvature and learning compensation terms in

the active contour stage.

Table 1: Evaluation of different conﬁgurations for our

method and comparison with GrabCut.

Method TP TN FP FN

SVM 65.22 95.42 4.58 34.78

Original AC 79.41 93.64 6.36 20.59

Original AC + SVM 84.2 89.91 10.09 15.80

Full Method 87.06 90.3 9.7 12.94

GrabCut 93.72 52.29 47.71 6.28

In a second set of experiments we compare our

method with GrabCut (Rother et al., 2004), a well-

known method for foreground extraction (see Ta-

ble 1). To outline the object for GrabCut we used the

external envelope of the parts identiﬁed by the per-

son detector. GrabCut has a good rate of true posi-

tives (i.e. good classiﬁcation of dress regions) but a

too high FP rate, i.e. tendency to classify background

as dress. This is likely due to the fact that GrabCut

performs better for extracting objects on an uniform

background, while out database contains many cases

where the background is complex or dress and back-

ground are visually similar (see Fig 4).

We also evaluate the segmentation by means of the

Jaccard (Intersection-Over-Union) score, score more

frequently used in papers dealing with image segmen-

tation:

S =

Surface(Y ∩Y

)

Surface(Y ∪Y

)

Also for this measure, our method (79.7%) largely

outperform GrabCut (64.86%).

Qualitative Evaluation. In this part we illustrate

the preceding conclusions with some examples taken

from the test database. A ﬁrst example of successful

segmentation was already shown in Fig. 2(e). In Fig. 4

we present some other difﬁcult examples of success-

Classiﬁcation-driven Active Contour for Dress Segmentation

(a) (b) (c) (d) (e)

(f) (g) (h) (i) (j)

Figure 6: More results on different types of background.

ful segmentation: (a,b) semi-transparent dress against

skin color and (c,d) black dress against cluttered back-

ground with people dressed in black. In Fig. 4(e, f) we

see a case of less successful segmentation: a red dress

on a red carpet. Here, the pixel descriptors are not

sufﬁciently discriminant to separate the foreground.

More examples are presented in Fig. 6 to illustrate the

behavior of the method with respect to different types

of background.

In Fig. 5 we show a visual comparison with Grab-

Cut: as hinted by the quantitative results, GrabCut

includes many background pixels in the foreground.

This occurs on all the database, explaining the high

FP rate in Table 1.

4 CONCLUSION

We presented a novel method for dress segmentation

that injects speciﬁc knowledge about the object into

a three stage detection model combining learning and

active contours. The inclusion of more training im-

ages, both for the person detector and for the proba-

bility map, together with the use of more sophisticated

pixel descriptors should allow to further improve the

results. In an extension of this work, we intend to

evaluate and adapt the method for other types of cloth-

ing items and, by replacing the person detector with

a deformable parts model (Felzenszwalb et al., 2010),

to other fashion objects.

REFERENCES

Chan, T. F. and Vese, L. A. (2001). Active contours with-

out edges. IEEE Transactions on Image Processing,

10(2):266–277.

Chang, C.-C. and Lin, C.-J. (2011). LIBSVM: A library

for support vector machines. ACM Transactions on

Intelligent Systems and Technology, 2:27:1–27:27.

Chen, H., Gallagher, A., and Girod, B. (2012). Describ-

ing clothing by semantic attributes. In Proceedings of

the 12th European Conference on Computer Vision -

Volume Part III, ECCV’12, pages 609–623.

Chen, Q., Huang, J., Feris, R., Brown, L. M., Dong, J., and

Yan, S. (2015). Deep domain adaptation for describ-

ing people based on ﬁne-grained clothing attributes.

In CVPR.

Datta, R., Joshi, D., Li, J., and Wang, J. (2008). Image

retrieval: Ideas, inﬂuences, and trends of the new age.

ACM Computing Surveys (CSUR), 40(2):1–60.

Deselaers, T., Keysers, D., and Ney, H. (2008). Features for

image retrieval: An experimental comparison. Infor-

mation Retrieval, 11(2):77–107.

Di, W., Wah, C., Bhardwaj, A., Piramuthu, R., and Sun-

daresan, N. (2013). Style ﬁnder: Fine-grained cloth-

ing style detection and retrieval. In Third IEEE In-

ternational Workshop on Mobile Vision, CVPR, pages

8–13.

Dong, J., Chen, Q., Shen, X., Yang, J., and Yan, S. (2014).

Towards uniﬁed human parsing and pose estimation.

In CVPR, pages 843–850.

Felzenszwalb, P. F., Girshick, R. B., McAllester, D., and

Ramanan, D. (2010). Object detection with discrim-

inatively trained part based models. IEEE Transac-

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

tions on Pattern Analysis and Machine Intelligence,

32(9):1627–1645.

Hsu, E., Paz, C., and Shen, S. (2011). Clothing image re-

trieval for smarter shopping (stanford project).

Jammalamadaka, N., Minocha, A., Singh, D., and Jawa-

har, C. (2013). Parsing clothes in unrestricted images.

In Proceedings of the British Machine Vision Confer-

ence.

Kalantidis, Y., Kennedy, L., and Li, L.-J. (2013). Get-

ting the look: Clothing recognition and segmentation

for automatic product suggestions in everyday photos.

In Proceedings of the 3rd ACM Conference on Inter-

national Conference on Multimedia Retrieval, ICMR

’13, pages 105–112.

King, I. and Lau, T. K. (1996). A feature-based image

retrieval database for the fashion, textile, and cloth-

ing industry in hong kong. In International Sym-

posium on Multi-Technology Information Processing

(ISMIP’96), Hsin-Chu, Taiwan.

Lew, M., Sebe, N., Djeraba, C., and Jain, R. (2006).

Content-based multimedia information retrieval: State

of the art and challenges. ACM Transactions on Mul-

timedia Computing, Communications, and Applica-

tions (TOMCCAP), 2(1):19.

Liu, S., Feng, J., Song, Z., Zhang, T., Lu, H., Xu, C., and

Yan, S. (2012a). Hi, magic closet, tell me what to

wear! In ACM Multimedia, pages 619–628. ACM.

Liu, S., Liang, X., Liu, l., Lu, K., Liang, L., and Yan, S.

(2014). Fashion parsing with video context. In ACM

Multimedia, pages 467–476. ACM.

Liu, S., Song, Z., Liu, G., Xu, C., Lu, H., and Yan, S.

(2012b). Street-to-shop: Cross-scenario clothing re-

trieval via parts alignment and auxiliary set. In CVPR,

pages 3330–3337.

Liu, Y., Zhang, D., Lu, G., and Ma, W.-Y. (2007). A sur-

vey of content-based image retrieval with high-level

semantics. Pattern Recognition, 40(1):262–282.

Nguyen, T. V., Liu, S., Ni, B., Tan, J., Rui, Y., and Yan, S.

(2012). Sense beauty via face, dressing, and/or voice.

In ACM Multimedia, pages 239–248.

Redi, M. (2013). Novel methods for semantic and aesthetic

multimedia retrieval. PhD thesis, Universite de Nice,

Sophia Antipolis.

Rother, C., Kolmogorov, V., and Blake, A. (2004). Grab-

Cut – interactive foreground extraction using iterated

graph cuts. ACM Transactions on Graphics, pages

309–314.

Scholkopf, B. and Smola, A. J. (2001). Learning with Ker-

nels: Support Vector Machines, Regularization, Opti-

mization, and Beyond. MIT Press.

Simo-Serra, E., Fidler, S., Moreno-Noguer, F., and Urtasun,

R. (2014). A high performance crf model for clothes

parsing. In ACCV.

Song, Z., Wang, M., sheng Hua, X., and Yan, S. (2011).

Predicting occupation via human clothing and con-

texts. In Proceedings of the 2011 International Con-

ference on Computer Vision, ICCV ’11, pages 1084–

1091, Washington, DC, USA. IEEE Computer Soci-

ety.

Yamaguchi, K. (Last checked: September 2015). Fashion-

ista image database, http://vision.is.tohoku.ac.jp/

˜kyamagu/research/clothing parsing/.

Yamaguchi, K., Hadi, K., Luis, E., and Tamara, L. B.

(2015a). Retrieving similar styles to parse clothing.

IEEE Transaction on PAMI, 37:1028–1040.

Yamaguchi, K., Kiapour, M. H., and Berg, T. L. (2013).

Paper doll parsing: Retrieving similar styles to parse

clothing items. In Proceedings of the 2013 IEEE Inter-

national Conference on Computer Vision, ICCV ’13,

pages 3519–3526, Washington, DC, USA.

Yamaguchi, K., Okatani, T., Sudo, K., Murasaki, K., and

Taniguchi, Y. (2015b). Mix and match: Joint model

for clothing and attribute recognition. In Xie, X.,

W. Jones, M., and K. L. Tam, G., editors, Proceedings

of the British Machine Vision Conference (BMVC),

pages 51.1–51.12. BMVA Press.

Yang, M. and Yu, K. (2011). Real-time clothing recogni-

tion in surveillance videos. In ICIP, pages 2937–2940.

IEEE.

Yang, Y. and Ramanan, D. (2013). Articulated human de-

tection with ﬂexible mixtures of parts. IEEE Transac-

tion on PAMI, 35:2878–2890.

Classiﬁcation-driven Active Contour for Dress Segmentation