Original article

Textile Research Journal

0(0) 1–30
The exploration of artificial intelligence ! The Author(s) 2021
Article reuse guidelines:
application in fashion trend forecasting sagepub.com/journals-permissions
DOI: 10.1177/00405175211006212
journals.sagepub.com/home/trj

Mengyun Shi1, Cali Chussid1 , Pinyi Yang2, Menglin Jia3,

Van Dyk Lewis1 and Wei Cao4

Abstract
Fashion trends today are changing much faster than ever before. Timely and reliable trend forecasting is, therefore,
critical in the fashion industry. Traditional fashion forecasting requires professionals to abstract image-based information
across design collections and time intervals from around the world, which is extremely time-consuming and labor
intensive. Considering the financial cost associated with manual labeling and the accuracy of classifications based upon
human subjective judgment, this explorative study proposes a data-driven quantitative abstracting approach using an
artificial intelligence (A.I.) algorithm. Firstly, an A.I. model was trained to be familiar with fashion images from a large-
scale dataset under different scenarios such as online stores and street snapshots; secondly, the model could detect
garments and classify clothing attributes such as fabric textures, garment style, and design details from runway photos
and videos; thirdly, the model could summarize fashion trends from the attributes it developed. The adoption of an A.I.
algorithm proved to be an objective and systematic computerized method of interpreting fashion dynamics in a more
efficient, accurate, sustainable, and cost-effective way.

Keywords
Fabrication, Fabrication, product, management of:, Systems, Product and Systems Engineering, brands, management of:,
Systems, Social Science, color, Chemistry, consumer behavior/demand, management of:, Systems, Social Science,
forecasting and business modeling, management of:, Systems, Social Science, artificial intelligence, fashion trend fore-
casting, fashion attribute detection

Detecting fashion attributes from images is critical for Sales and marketing executives abstract information
professionals in the fashion industry, especially those in across product lines each season to recognize selling
trend forecasting. Fashion forecasting firms such as points. Fashion journalists and bloggers abstract infor-
WGSN, for example, detect fashion attribute informa- mation across runway photos to deliver symbolic core
tion from fashion images around the world ranging concepts that can be translated into editorial features.3
from fashion shows, visual merchandising, blogs, and Curators from museums and libraries abstract across
streetwear.1 They examine the information through historical fashion photos and collections to archive
experiences, observations, media scans, interviews, them in digital formats. Fashion scholars abstract
and exposure to new places. This type of information
analyzing process is known as abstracting: recognizing
similarities or differences across garments and collec- 1
Department of Fiber Science & Apparel Design, Cornell University, USA
tions.2 Many professionals in the fashion industry, 2
Department of Statistics, University of California, Los Angeles, USA
spanning a wide variety of roles and functions, utilize 3
Department of Information Science, Cornell University, USA
fashion abstracting in different ways. Fashion forecast- 4
Department of Family and Consumer Sciences, California State
ers abstract information across design collections and University, USA
time intervals to identify changes in fashion trends.
Corresponding author:
Designers, product developers, and buyers abstract Mengyun Shi, T57 Human Ecology Building, Cornell University, Ithaca,
information across garment groups or collections NY14853-4401 USA.
to develop cohesive and visually appealing lines. Email: [email protected]
2 Textile Research Journal 0(0)

both historical and modern fashion photography for September. Throughout that month, top fashion edi-
their own research purposes. tors from around the world gather together to observe
Analyzing image-based fashion attributes is, howev- a plethora of designers’ collections. After the fashion
er, time-consuming and labor intensive, even for pro- shows are over, it is their job to absorb what they saw
fessionals. WGSN, for example, regularly employs in each show and try to pull out the major trends.
around 150–200 forecasters to analyze the trends for Although there are runway images that editors can
each season.1 Furthermore, trends nowadays are reference back and forth, it is very tedious and chal-
changing much faster than before. Timely and reliable lenging to single-out overarching trends. It is especial-
trend forecasting is becoming much more important in ly difficult when considering that these editors have to
order to meet consumers’ needs. Manual labeling and pull trends for a wide range of dimensions including,
labor-intensive classification of visual information limit but not limited to, fabric detail, fabric color and gar-
the speed and reliability of interpreting fashion dynam- ment silhouette. While researching various fashion
ics and spreading them all over the world due to the
magazines, one notices that trend predictions for the
amount of work required. In addition, manual labeling
same season can be completely different even if the
based on an individual’s experience is quite subjective
editors attended the exact same fashion shows. This is
due to human errors or opinions on certain fashion
due to the fact that the prediction is subjective and
attributes. A possible breakthrough would be taking
advantage of widely used artificial intelligence (A.I.) mainly based on editors’ personal experiences and
methods that have already been adopted in fields out- expertise.
side of fashion. These technologies could improve fash- Another disadvantage in fashion editors’ trend anal-
ion forecasters’ analysis and classification of visual ysis is that their predictions are limited to the collec-
information by providing objective and systematic tions being shown on the runways during that
computerized methods. particular season only. This may raise the concern
Considering the financial cost associated with manual that the information from fashion show collections
labeling and the accuracy of classifications and attribute can be limited and may not represent what people
detection based upon human subjective judgment, this wear in real life. Fashion editors utilize the trickle-
explorative study proposes a data-driven, quantitative down approach assuming that the trends presented in
abstracting approach using an A.I. algorithm. The effec- runways will diffuse down to lower echelons, but this is
tiveness of the A.I approach was compared with tradition- not always the case. The trickle-up theory, on the other
al methods. A.I. can be defined as the ability of computers hand, proves that street and subculture trends can actu-
and machines to perform tasks usually designated to ally influence luxury fashion. Hence, the A.I. algo-
human beings. The A.I. algorithm used in this study is rithm’s trend analysis is much more holistic because it
an augmented version of an object detection algorithm captures and abstracts across fashion images on both
called “Faster R-CNN” developed by A. I. researchers ends of the spectrum ranging from fast fashion to
from Facebook.4 This algorithm has been applied in the runway fashion.
field of self-driving cars for detecting pedestrians and traf- Previous work related to our proposed study can be
fic signs. There is no research documenting the application divided into four main sections: (1) fashion attribute
of this algorithm in detecting fashion attributes from fash- and style prediction; (2) fashion trend analysis and
ion images and analyzing fashion trends. data mining; (3) fashion datasets; and (4) fashion
The algorithm could potentially improve the effi- image recognition and retrieval from different
ciency and lower the cost of analyzing fashion images domains. The fashion attribute and style prediction
through the following methods: firstly, an A.I. model
section covered the existing A.I. algorithms that have
was trained to recognize images from a large-scale
been used for detecting fashion attributes or styles
dataset under various scenarios such as online stores
directly from fashion images. The fashion trend analy-
and street snapshots; secondly, the model was used to
sis and data mining section included the current appli-
detect garment details and classify clothing attributes
cation of A.I. algorithms for fashion trend forecasting
such as fabric textures, garment style, and design
details from runway photos and videos; thirdly, the and data mining-related tasks. For the fashion datasets
model was used to summarize fashion trends based section, a variety of fashion datasets built for training
on the attributes it detected. A.I. models were reviewed. In the fashion image recog-
nition and retrieval from different domains section,
previous research work that used different domains
Literature review instead of the reviewed datasets to train the A.I. algo-
In the traditional fashion cycle, a designer’s Spring/ rithms for fashion attributes or image retrieval were
Summer collection is shown on the runway in discussed.
Shi et al. 3

Detecting clothing categories and attributes from interpreting fashion images originating from separate
fashion images domains. For example, Jia et al. (2018) utilized an A.I.
model trained from a dataset of images across different
Fashion attribute detection has received increasing domains such as e-commerce photos and street style
attention from computer vision communities in recent images to predict both runway show images and fash-
years.5 Fashion-related papers using the proven scien- ion technical sketches.25 Another example of transfer
tific method increased by 20% from 2010 to 2017 at top learning is training a model to gather human-annotated
computer vision conferences Computer Vision and fashion attributes from a variety of sources on the web
Pattern Recognition (CVPR) and International and learn to predict fashion attributes of images from
Conference on Computer Vision (ICCV). This indi- different sources.26,27
cates a steady traction in this area.6 There has also The fashion style prediction phase has been tackled
been a large body of research involving clothing model- by researchers using two distinct methods: supervised
ing, attribute recognition, image parsing, image retriev- learning and unsupervised learning. Supervised learn-
al, and styling recommendations. Image retrieval in ing entails teaching A.I. models to recognize fashion
particular has gained interest lately since it is an essen- styles based on training data manually annotated by
tial task for consumers in fashion e-commerce. The human experts while unsupervised learning utilizes
majority of the research focuses on retrieving daily models, representations, and frameworks that learn
life images and online shopping ones.7,8 Currently, style recognition models without using any human-
only two articles target runway images, but neither annotated training data. Researchers have used super-
detects fashion attributes from fashion industry profes- vised learning by obtaining a small dataset, defining
sionals’ perspectives. Fashion attribute detection for parameters, and using A.I. models to extract features
“fashion insiders” requires a more objective, detailed, and style attributes from different sets of data.28 Other
and in-depth analysis of fashion images than that for researchers have used supervised learning in order to
regular consumers. construct fashion semantic spaces that can be used to
Earlier works relied on some classical computer describe a wide span of fashion styles.29–31 In the unsu-
vision approaches to sort apparel into categories and pervised learning space, researchers32 have used poly-
describe fashion attributes from images.5,9 –11 Further lingual topic modeling to better understand how
research was dedicated to a computer vision technique low-level attributes such as color and silhouettes translate
called “image segmentation” for classifying different to high-level styles such as “Bohemian.” In the social
apparel categories via probabilistic methods.12–14 In network realm, A.I. models have also been explored by
order to improve the accuracy in recognizing garments, researchers in the unsupervised learning space to learn
some of the research mentioned above contained pre- style attributes of outfits by using a large collection of
processed images to account for human posture and unlabeled user-created style images.33,34
body parts (Figure 1(a)). In addition, some computer vision researchers
Gu et al. (2020) identified clothing attribute predic- focused on retrieving images that have high similarity
tion as a major component of the fashion attribute pre- in fashion attributes.8,28 Other researchers began to use
diction phase that can be denoted by three different more advanced A.I. approaches that demonstrated sig-
types of learning16: (1) single-task learning; (2) multi- nificant boosts to the accuracy of summarizing fashion
task learning; and (3) transfer learning. The simplest, attributes. For example, many researchers started to
single-task learning, involves identifying clothing utilize an A.I. algorithm called “R-CNN” to detect gar-
attributes from specific fashion domains. For example, ment parts or generate attribute descriptions of fashion
Di et al. (2013) explored clothing style detection by first images (Figure 1(b)).26,27,35
compiling human annotations of a list of clothing There are three main differences between previous
attributes used to train an A.I. model to recognize studies and our research. Firstly, we developed a new
and retrieve different clothing styles.17 Other research- A.I. model architecture that enables us to quantify
ers have acquired attribute recognition models by gath- attributes detected from fashion images into five cate-
ering weakly labeled data from e-commerce sites.5,18–20 gories (textures, fabrics, shapes, parts, style) which can
Multi-task learning has the ability to perform multiple facilitate the post-analysis on these attributes for trend
tasks such as learning clothing attributes and landmark forecasting tasks and produce more meaningful
detection at the same time. In this area, researchers insights to fashion professionals in a more quantitative
have introduced A.I. models that can jointly recognize way. In contrast, previous models developed in the
clothing and attributes.21 Other applications of multi- above studies can only predict fashion attributes with-
task learning involve simultaneously analyzing human out telling which specific categories these attributes
body poses and attributes.22–24 Lastly, transfer learning belong to. Secondly, unlike the previous approaches
recognizes clothing attributes by examining and used to produce fashion attribute prediction and style
4 Textile Research Journal 0(0)

Figure 1. (a,b) Previous methods used to summarize clothing categories and attributes from fashion images.13,15

prediction in other A.I. models, our proposed professionals in the fashion industry who know what
approach can unify both prediction tasks and detect specific attributes and depth of knowledge are most
fashion attributes and style simultaneously in a single important to fashion insiders. The method proposed
A.I. model which is a more efficient way for trend fore- in this study is mostly based on an augmented version
casting tasks. Thirdly, most of the aforementioned of an A.I. algorithm called “Faster R-CNN.”4
research has been done by computer scientists, not Compared to all previous methods, the Faster
Shi et al. 5

R-CNN algorithm has been proven to provide accurate knowledge, no similar studies have been previously
detection results in images.36 conducted.

Analyzing fashion trends based on detected fashion Fashion datasets

attributes and data mining There are many different types of benchmark datasets
Fashion data mining involves extracting insights from available to fashion researchers to leverage for their
a wide variety of fashion data. It has become an studies. Single-task datasets are meant to focus on
extremely promising tool for the fashion industry and one particular task such as style prediction, while
is expected to grow in the future. The two main focus multiple-task datasets are able to combine multiple
areas of this research are fashion trend analysis and tasks at the same time. An example of a single-task
hybrid analytics. In the fashion trend analysis realm, data set is the RunwaytoRealway dataset,8 which con-
researchers have explored runway and real world fash- sists of over 345,000 images that can be used for
ion and use computer vision to provide intrinsic and analyzing visual trends in fashion and span over the
extrinsic evaluations to understand outfit similarity, course of 15 years. Another example of a single-task
etc.8 Other applications of fashion trend analysis have benchmark dataset is the Hipster dataset, which focuses
involved focusing on and analyzing street style fashion on defining key styles such as “goth” and
through visual embedding of garment styles.37–39 On “preppy.”14,28,43–45 The DeepFashion dataset15 is an
the other hand, researchers specializing in hybrid ana- example of a multi-task dataset. It contains over
lytics have used A.I. technologies to explore how cloth- 800,000 images that can achieve multiple tasks includ-
ing features relate to other factors such as one’s ing landmark detection, clothing retrieval, etc.
occupation or urban tribe.40–42 FashionAI is another multiple-task benchmark dataset
In addition, the previous work on trend analysis that contains extensive information related to land-
broke down catwalk images from New York fashion mark detection and attribute prediction covering a
shows to locate style trends in high-end fashion.7 The wide variety of categories and subcategories in
recent advance in A.I. algorithms enabled more work women’s clothing.28,35,46–48
and advancements in this area. Several recent In this study, we adopted the DeepFashion dataset
approaches utilized an A.I. algorithm to extract cloth- for two main reasons: (1) the images from the
ing attributes from images and create a visual embed- DeepFashion dataset are taken under different scenari-
os including stores, street snapshots, and consumers
ding of clothing style clusters to identify fashion trends
(user-taken photos). Furthermore, the associated anno-
in clothing around the globe.12,37,43 However, it is not
tations (meta-data) of the collected images from
quite clear whether the trends discovered in these stud-
DeepFashion are labeled by various people and sources.
ies by computer scientists are informative enough or
With such diverse image types and annotations, it ena-
not and whether the trends developed were consistent
bles our A.I. model to learn the fashion attributes of a
with the results reported by fashion professionals
certain type of clothing item from diverse perspectives
worldwide.
which makes our model prediction more objective and
Furthermore, most previous studies conducted by
less biased. (2) The fashion attributes annotated in
researchers in the computer science domain are more
DeepFashion images are categorized into five groups
in the field of consumer-facing applications of A.I.
characterizing texture, fabric, shape, part, and style,
models, not fashion professional-facing applications.
respectively. The structure of this dataset is aligned
This is possibly due to the fact that computer science
with the A.I. model architecture we developed in this
researchers lack background or experience in the fash-
study which helps our A.I. model learn to recognize
ion industry. In this study, we focus on fashion
fashion attributes from five different perspectives and
professional-facing applications of A.I. models and
in a more systematic way.
investigate how A.I. technology can be adopted to
help fashion professionals analyze trends in a more
efficient way. Specifically, we investigated the two
Fashion images recognition and retrieval from
areas as follows: (1) compared the fashion trends sum- different domains
marized by the proposed A.I. algorithm with predic- There have been a number of research works tackling
tions from the world’s leading fashion magazines (such the issue of cross-domain fashion image retrieval and
as Vogue, ELLE Magazine, and Harper’s Bazaar); and recognition. The most popular topic in this area is the
(2) compared the fashion trends summarized by the retrieval of similar fashion images from different
proposed A.I. algorithm of ZARA’s products with domains such as streetwear and online shopping web-
forecasting analyses prepared by Vogue. To our sites. This is extremely useful for e-commerce fashion
6 Textile Research Journal 0(0)

marketing.16,38,49 Most of the work in retrieving images 2.2: is there any difference between the fashion
is based on the current advance of the A.I. approach in trends summarized by the proposed A.I. algo-
which it is necessary to teach the A.I. algorithm to rithm of ZARA’s products and the forecasting
recognize attributes shared by similar fashion items analyses made by Vogue?
from both streetwear and online shopping
domains.27,35,50,51 Industry background and proposed
All the research mentioned above is conducted by
computer scientists and/or engineers who lack knowl-
methodology
edge and experience in the fashion industry. Detecting fashion attributes using A.I.
Specifically, they did not consider the application of
the A.I. model in generating fashion attributes from In order to use the proposed A.I. method in analyzing
static images and predicting fashion trends from fash- fashion images, the A.I. algorithm had to go through a
ion insiders’ perspectives. To our knowledge, this process called “training.” During the training process,
research is the very first study to: (1) evaluate the accu- the A.I. algorithm was taught the concept of fashion
racy of the fashion attributes generated by the A.I. attributes from an image dataset annotated with cer-
model through analyzing the Dolce & Gabbana tain fashion attributes (Figure 2(b)). After that, the A.I.
Spring 2018 ready-to-wear collection sample and addi- algorithm started to identify clothing attributes in new
tional Spring/Summer 2018 designer runway images; images. For instance, the A.I. algorithm learned the
(2) compare the effectiveness of trend forecasting devel- concept of “maxi dress” from the annotated dataset
oped by the A.I. model with the world’s leading fashion and then predicted the “maxi dress” attribute in mil-
journals through examining around 10,000 ZARA lions of unlabeled images thereafter. In this study, a
large-scale fashion dataset called “DeepFashion” was
women’s product images in Summer 2018; and (3)
adopted. It contains over 280,000 fashion images gath-
explore the possibilities of summarizing fashion attrib-
ered from both online shopping websites and street
utes from moving model images in fashion videos by
photos with relatively comprehensive annotations of
the A.I. model. The overall goal of the study is to pro-
garment details.15 Meanwhile, a powerful A.I. algo-
vide a more scientific and efficient trend forecasting
rithm called “Faster R-CNN” was also utilized to
solution for fashion insiders using an A.I. model.
leverage the DeepFashion dataset.4 The A.I. algorithm
was used to analyze simple random sampling of
A.I. augmented fashion trend analysis conceptual 16 looks in Dolce & Gabbana’s Spring 2018 ready-
framework to-wear collection and then to examine other Spring/
A fashion trend analysis conceptual framework using Summer 2018 designer runway show images according-
the A.I. method was proposed (Figure 2(a)). The ly. The testing results are interpreted in a later section.
framework is divided into two stages: (1) detecting
fashion attributes using A.I. from static runway Details of the training dataset. Diverse image domain,
images and moving fashion models in runway videos; source, and pairs. According to the DeepFashion
and (2) analyzing fashion trends based on fashion research,15 the images used for training our A.I.
attributes detected in Stage 1. The evaluation methods models were taken under different scenarios – from
were also provided at each stage to test the effectiveness stores to street snapshots to consumers (user-taken
photos). Overall, 44.2% of the images contained cloth-
of the results by the proposed A.I. system. (Note: The
ing items having four to five images of varied poses and
results from tracking moving models will be examined
viewpoints; 13.1% of the images consisted of several
in future studies.)
user-taken photos of exactly the same clothing item.
The developed research questions are as follows:
Furthermore, the images from DeepFashion are not
1. Will the proposed A.I. model detect considerably only collected from leading online shopping websites
accurate fashion attributes from: such as Forever21, but also from the Google image
1.1: static runway images? search engine. This guarantees that the images come
1.2: moving fashion models in runway videos? from various sources such as blogs, forums and other
2. When using the A.I. model to analyze fashion user-generated content which extends beyond the
trends based on the developed fashion attributes, image datasets collected solely from shopping websites.
2.1: is there any difference between the fashion Such a diverse image domain and source enabled
trends summarized by the proposed A.I. algo- our proposed A.I. model to learn the fashion
rithm and the predictions from the world’s attributes of a certain type of clothing item from
leading fashion magazines (such as Vogue, diverse perspectives which ensured our model predic-
ELLE Magazine, and Harper’s Bazaar)? tion was more objective and less biased. In contrast,
Shi et al. 7

Figure 2. (a–c) The A.I. augmented fashion trends analysis methods.

8 Textile Research Journal 0(0)

if the model was trained on a single image domain, it branches.58 We trained the modified Faster R-CNN
would only be able to recognize attributes well if they to learn both fashion attributes and garment categories
belonged to the trained image domain and result in as a multi-task learning problem. A more technical
severe bias. explanation of the model architecture that we devel-
Diverse annotations. The associated annotations oped in this study can be found in the supplementary
(meta-data) of the collected images from material section.58
DeepFashion are labeled by various people and sour-
ces. The human annotators are also recruited to filter
out images with low quality annotations. This further
helped our A.I. model learn how to recognize fashion Analyzing fashion show videos. This study further evaluat-
attributes in a less biased way. In contrast, the trend ed the effectiveness of the proposed A.I. system in iden-
reports made by leading magazines are usually written tifying fashion attributes from analyzing moving
by a single editor.52–55 This could lead to biased pre- models in runway videos which is much more challeng-
dictions due to each fashion editor interpreting trends ing than static fashion images. This is because runway
according to his/her own preferred attribute/style videos consist of more complicated content and scenes
preference. in a moving format which increases the difficulty of
Dataset modification. There are some labeling prob- analysis. Furthermore, to summarize the information
lems in the dataset we chose, DeepFashion (category in runway videos requires the proposed A.I. system to
and attribute prediction benchmark). For the attributes track moving fashion models in videos frame-by-frame.
annotated in the DeepFashion dataset, we manually
In order to prevent the proposed A.I. system from
removed 45 unclear attributes (such as “girl” and
losing track of moving fashion models in runway
“please”) and merged semantically similar attributes
videos, a technique called “image segmentation” was
(for example, “abstract geo” vs. “abstract geo print”
adopted in this study. Image segmentation is a typical
vs “geo” vs “geo pattern” vs “geo print”). The resulting
technique to locate objects and boundaries (lines,
dataset contained 544 diverse clothing attributes and 50
clothing categories. Wrongly labeled (false positive) curves, etc.) in images automatically used in the com-
categories, attributes, and bounding boxes may still puter vision field. Figure 3 shows one example of the
exist in this dataset (e.g., recognizing a skirt as a segmentation technique. It was used to locate the out-
dress), but we were not able to deal with all of them lines of the person, the sheep, and the dog in the image
due to the expensive human labor costs involved in automatically.59
correcting wrong labels for a large amounts of images. The video of the Burberry Prorsum Fall 2012
Menswear Collection was selected randomly to evalu-
Details of the proposed A.I. model. We extended the Faster ate the effectiveness of the proposed A.I. system in
R-CNN object detection framework with ResNet 10156 detecting fashion attributes from tracing moving fash-
and ROI-align4 (implemented by Google Research36 ion models.
based on Tensorflow Deep Learning library57) with For the video analysis, we used the Mask R-CNN
two modifications: a pruning mechanism and addition- models developed by the Facebook Artificial
al clothing attribute branches parallel to category Intelligence Group.60

Figure 3. (a) the original image; (b) “image segmentation” technique is used to locate the outlines of the person, sheep, and dog in
the image.59
Shi et al. 9

Analyzing fashion trends Around 10,000 ZARA product images from the
Spring/Summer season were used in our analysis.
The goal of this stage was to identify fashion trends in a
Similar to the analysis of the runway images, the
certain season by summarizing the fashion attributes
python programming language64 and the pandas
detected from the previous stage. In order to do so,
library65 were used to calculate the frequency of
we adopted the descriptive statistics method of summa-
detected attributes for each category. The matplotlib
rizing fashion trends (Figure 2(c)). To further test the
library66 was used to visualize the calculated frequency
effectiveness of the proposed framework at this stage,
of each attribute.
we investigated the two areas as follows: (1) compared
the difference between the trend analysis generated by
the proposed A.I. model and the forecasting by the Results and discussion
fashion editors from leading magazines such as
Vogue, ELLE Magazine, and Harper’s Bazaar; and The result of detecting fashion attributes
(2) compared the difference between the trend analysis From static runway images. To answer Research question
generated by the proposed A.I. model based on 1.1, simple random sampling of 16 looks in Dolce &
ZARA’s products and the forecasting by Vogue. Gabbana’s Spring 2018 ready-to-wear collection were
selected. Figure 4 presents the rich attribute descrip-
Comparing the fashion trends analyzed from the A.I. algorithm tions generated by the A.I. model in detected regions.
with the predictions from the world’s leading fashion The model correctly identified eight looks belonging to
magazines. The Spring/Summer 2018 runway trend the dress category with high confidence (higher than
forecasting report was selected to compare with the 85%). The styles included rose, summer, cute, and
results developed by our A.I. model. The trend fore- party in agreement with the lifestyle called “the sweet
casts made by fashion editors from Vogue, ELLE life” which the brand constantly celebrates. It accurate-
Magazine, and Harper’s Bazaaar were presented in ly identified six floral patterns (one with lace), five pat-
New York, London, Milan, and Paris.54,55,61–63 The terns as abstract, and the shapes of the dresses as either
way in which fashion editors collect trend information maxi or midi. The model predicted one dress as
and predict trends is very different from how an A.I. both maxi and midi with the same confidence (75%)
model does. It is important to understand the distinc- (Figure 4-1) which was mainly due to the length of the
tion since it may account for the slight differences in dress in between the knee and ankle (Figure 4-1). The
results outputted by the respective modes of model also identified two dresses that belonged to the
forecasting. bodycon silhouette (Figure 4-2 and Figure 4-6). This
In our analysis, we utilized 1844 runway images result is very promising since this designer is known for
spanning 46 collections from the Spring/Summer hourglass shapes and corseted tailoring.
season presented in Milan, London, New York, and The A.I. model, however, was not perfect. Figure 5
Paris. For the descriptive statistical analysis, the features some of the detected images with predictions
python programming language64 and the pandas that were not fully correct. The model behaved poorly
library65 were used to calculate the frequency of in detecting blazers, trousers, and skirts. For example,
detected attributes, and the matplotlib library66 was the pantsuit in Figure 5-2 was incorrectly attributed as
used to visualize the calculated frequency of each a dress, sweater, leggings, and sweatpants. The A.I.
attribute for each category. model was also confused by silhouettes containing sim-
ilar shapes. For example, the front opening of the
Comparing the fashion trends summarized from the A.I. blazer in Figure 5-2 was wrongly classified as a V-
algorithm based on ZARA’s products with the forecasting by neck due to the similar V-shapes.
Vogue. Women’s products from ZARA.com were While analyzing the accuracy of fashion attributes
selected for the study since ZARA is recognized as built from the sample data (16 looks in the Dolce &
the world’s largest apparel retailer. Analyzing fashion Gabbana runway show images) (Figure 6), a rate of
attributes from existing product images at ZARA pro- correct predictions made by the A.I. model (which in
vides an initial representation of current fashion trends short is called accuracy rate) was generated for each
around the world. The A.I. algorithm developed at look. Assuming that the mean of the accuracy rates
Stage 1 was used to analyze all ZARA women’s prod- in the population would be equal to or greater than
uct images (around 10,000 products). Then, the 75%, we performed the test on the random sample
detected fashion attributes were regrouped and summa- (16 looks) and set the significance level at 0.05. The
rized using descriptive statistics to abstract fashion p-value at 0.3827 (greater than 0.05) indicated that
trends. the mean for the population (Summer/Spring 2018
10 Textile Research Journal 0(0)

Figure 4. Example images using the proposed A.I. model. Each number after an attribute represents how confident the model’s
predictions are.

designer runway fashion show images) is equal to or fashion attribute descriptions from detected regions in
greater than 75%. the images. The variable which may affect the predic-
In this study, 75% correct descriptions per image tion accuracy made under the current observation was
was defined as “accurate.” Based on the resulting garment types, with an accuracy rate lower than 75%.
rates shown in Figure 6, we can conclude that the pro- To achieve a higher accuracy rate in the model, it might
posed A.I. model can generate considerably accurate be helpful to include more sample data to see how each
Shi et al. 11

Figure 5. Example images using the proposed A.I. model. Each number after an attribute represents how confident the model’s
predictions are. The attributes highlighted in red represent incorrect predictions.

variable such as garment type may influence the results The A.I. model accurately tracked and bound the
in a future study. male fashion models and their garments in the images
even though the fashion models were moving quickly.
From moving fashion models in runway videos. Figures 7 It also identified accessories such as ties, handbags, and
and 8 present some results using segmentation techni- umbrellas carried by the fashion models. However, The
ques to track the fashion models in the Burberry A.I. model had trouble recognizing other garments
Prorsum Fall 2012 Menswear fashion show. such as blazers, coats, and pants. Furthermore, the
12 Textile Research Journal 0(0)

Figure 6. The results of fashion attribute summaries produced by the proposed A.I. model in the Dolce & Gabbana Spring 2018
runway show.

tracking quality of the fashion models became ability to analyze all runway images and provide a
weaker when the fashion models walked in a row at more comprehensive overview of all possible fashion
the end of the show (Figure 9). One possible explana- trends in a much more efficient and encompassing
tion is that the overlap of these fashion models con- way. This type of analysis would be very challenging
fused the A.I. model, which increased the challenge to for any fashion editor due to the amount of work and
distinctly recognize each fashion model and his respec- time required. Another possibility is the discrepancy
tive garments. between how a fashion editor names a trend and how
the A.I. model labels that same trend. For example,
The result of trend analysis across 1844 runway images spanning 46 collections of
The comparison between the fashion trends summarized from the Spring/Summer season presented in Milan,
the A.I. algorithm and the predictions from the world’s leading London, New York, and Paris, our A.I. model identi-
fashion magazines. Before trend analysis, the A. I. fied the texture attribute “printed” among 1023 runway
Model was used to summarize more fashion attributes images. Comparatively, Vogue Italy and Vogue France
from the Spring/Summer 2018 designer runway show also indicated “printed” as a popular attribute.
images (Figure 10). However, the other top magazines did not mention
Figures 11 to 16 provide a more detailed glimpse “printed” in their prediction, which means either they
into the various attributes captured by the proposed used a different term to describe it or they omitted this
A.I. model including texture, fabrics, shapes, parts, trend entirely.
style, and category. The attributes in each respective Additional discrepancies between the A.I. model
graph with the highest frequency on the y-axis are the results and fashion publications were found when com-
ones that the A.I. model extrapolated as the most prev- paring other attribute categories. For example, in the
alent trends in the Summer/Spring 2018 season based fabric category, denim was identified as a trend in the
on the runway images. A.I. model analysis as well as in all major magazines
In addition, to test the effectiveness of the A.I except for Vogue Italy. In shape attributes, the A.I.
approach, the results captured by the proposed A.I. model predicted a trend called “pencil” that was not
model from analyzing runway show images were identified by editors in the majority of the fashion mag-
compared to the trends reported in leading magazines azines. These discrepancies are most likely due to the
such as Vogue, ELLE Magazine, and Harper’s different naming conventions among fashion maga-
Bazaar.54,55,61–63 We found fashion editors missed a zines and the A.I. model.
lot of attributes that our A.I. model extrapolated. Lastly, there were some attributes such as parts and
The possible reason is that the A.I. model has the styles mentioned in magazines but not identified by the
Shi et al. 13

Figure 7. Tracking fashion models using “image segmentation” technique - 1. The upper row is the original video frame. The lower
row is the result using segmentation technique. The video adapted from Burberry Fall 2012 Menswear Runway Recap. (n.d.).
Retrieved 24 September 2018 from www.youtube.com/watch?v=w2YCUbf86zA.

A.I. model. For instance, trimming was recognized as a attributes such as “bold color” can encompass a wide
major trend among all magazines as well as the A.I. range of colors.
model. However, the A.I. model did not capture some
of the micro trends spotted by editors in this category The comparison between the fashion trends summarized from
such as “puffed” and “padded shoulders.” There was the A.I. algorithm based on ZARA’s products and the
also a big disagreement with regard to the most popu- forecasting by Vogue
lar color(s) of the season. This might be due to the fact Garment style and detail trends. In addition to fash-
that colors are subjective for each editor and some ion show imagery, ZARA fashion images were selected
14 Textile Research Journal 0(0)

Figure 8. Tracking fashion models using “image segmentation” technique - 2. The upper row is the original video frame. The lower
row is the result using segmentation technique. The video adapted from Burberry Fall 2012 Menswear Runway Recap. (n.d.).
Retrieved 24 September 2018 from www.youtube.com/watch?v=w2YCUbf86zA.

to be analyzed by the A.I. model based on its leading detected attributes (Figure 23(a) and (b)). These results
position in the world’s apparel retailing industry. The can be viewed as seasonal trends since it is possible that
detailed analysis of fashion images at ZARA could ZARA might intend to sell more appropriate summer
reveal whether ZARA aligned with the trend forecasts clothes such as short sleeves versus long sleeves during
of a certain season such as Summer 2018 to verify the this season.
effectiveness of the trend analyses made by the A.I. To further verify the accuracy of trend analysis by the
model. Figure 23 presents some of the developed gar- proposed A.I. model, Summer 2018 fashion trends from
ment style and detail trends from ZARA images. Tops, Vogue were used as a secondary source of information for
T-shirts, and dresses represented the highest frequency comparison.52 The results showed that both Vogue and
in the “Garment style” attributes with styles such as the A.I. model listed oversized, floral, shirtdress, and
round neck and short sleeves being the most frequently polka dot as the popular fashion trends for Summer
Shi et al. 15

Figure 9. Tracking fashion models using “image segmentation” technique - 3. The upper row is the original video frame. The lower
row is the result using segmentation technique. The video adapted from Burberry Fall 2012 Menswear Runway Recap. (n.d.).
Retrieved 24 September 2018 from www.youtube.com/watch?v=w2YCUbf86zA.

2018. However, Vogue also listed additional summer down to product lines in the fast fashion and street style
trends including baby doll, lightweight, check & plaid, realms. For example, ZARA may not have included the
and patterns not found in the trend report produced by baby doll silhouette in its women’s product offering
the A.I. model. Possible explanations for these inconsis- simply because this style is very dramatic and more
tencies include: (1) “Check & plaid” were never labeled in runway-esque, not for consumers’ daily wear.
the large dataset “DeepFashion” images where the A.I.
model was trained before detection; therefore, there was Color trends. Figure 24(a) presents some color trend
no way for the model to pick up on those trends and analysis results produced by the A.I. model. White and
name them explicitly during the analysis. (2) ZARA did black occupied the top two positions in the “Color
not include these attributes in their mood board while trends” plot. However, it is safe to assume that these
designing their product lines, which made it impossible are not particular Summer 2018 trends since consumers
for the A.I. model to recognize these attributes. This usually wear these two colors across all seasons, not
might be due to the fact that fashion editors prepare only for Summer 2018. Trendy colors typically vary
trend predictions based solely on runway show images over seasons.
and not all high-end fashion on the runway will translate
16 Textile Research Journal 0(0)

Figure 10. More results of fashion attribute summaries in the Spring/Summer 2018 runway fashion show.

Summer 2018 color trend analysis in Vogue was results exhibited some interesting, but less obvious,
adopted as a secondary source of information to eval- connections to Vogue’s insights. For instance, Vogue
uate the accuracy of the color trend forecasting ana- suggested the following colors as the leading trends
lyzed by the proposed A.I. model.53 The comparison for Summer 2018: sky-blue, tomato red, light green,
Shi et al. 17

Figure 11. The result of detected textures in the Spring/Summer 2018 runway shows. The y-axis indicates how many times the
attributes have been detected by our A.I. model on runway shows Spring/Summer 2018.

and military green. Findings from the proposed A.I. Fashion editors predict color trends by noting the top
model indicated that blue, red, and green were discov- colors presented across different runway shows and
ered as the most popular colors, but it did not specify then scroll through Pantone’s color palettes to give
which specific shades of each represented the color these colors specific names. The A.I. model extracts
trends for Summer 2018. Therefore, further investiga- generic color names based on what DeepFashion
tion on the color attributes from ZARA products was images are fed into the model and are not the same
conducted. It was found that ZARA did include the as Pantone swatch names.
sky-blue color in 98 out of their 10,000 products The results of fiber content analysis in garment fab-
(Figure 24(a)), but it is hard to say whether ZARA rics from ZARA are also shown in Figure 24(b). The
considers sky-blue as a trendy color. For the red study did not further evaluate the accuracy of these
color, it was confirmed that ZARA did not produce results, since Vogue failed to provide any fabric infor-
any collection with tomato red, but some products mation in their Summer 2018 trend report.
were a mix of white and red (47 products). For the
green color, ZARA did not design any products asso-
ciated with light green or military green. We believe the
Discussion
discrepancies noted above are due to disparate infor- Research question 1.1 (Will the proposed A.I. model
mation between Vogue and ZARA rather than a mis- detect considerably accurate fashion attributes from
take made by the A.I. model. It is also understandable static runway images?) was supported by the fact that
that the A.I. model did not predict the colors exactly the proposed A.I. model can generate accurate fashion
the same as Vogue since they are very specific and attribute descriptions from detected regions in the
nuanced shade names. Other explanations derive images with the average precision of around 75%.
from the nature of how color trends are developed. Research question 1.2 (Will the proposed A.I. model
18 Textile Research Journal 0(0)

Figure 12. The result of the detected fabrics in the Spring/Summer 2018 runway shows.

Figure 13. The result of the detected shapes in the Spring/Summer 2018 runway shows.

detect considerably accurate fashion attributes from the predictions from the world’s leading fashion maga-
moving fashion models in runway videos?) was somewhat zines such as Vogue, ELLE Magazine, and Harper’s
established since the A.I. model was able to track and Bazaar?) was partially answered since our A.I. model
bind the male model with moving images, summarizing discovered some fashion attributes or trends that fash-
some garments and accessories successfully. ion editors missed, possibly due to the fact that the A.I.
Research question 2.1 (When using the A.I. model to model has the ability to provide a more comprehensive
analyze fashion trends based on the developed fashion overview of all possible fashion trends in a much more
attributes, is there any difference between the fashion efficient and encompassing way by analyzing all
trends summarized by the proposed A.I. algorithm and runway images. Meanwhile, the A.I. model failed to
Shi et al. 19

Figure 14. The result of the detected parts in the Spring/Summer 2018 runway shows.

Figure 15. The result of the detected styles in the Spring/Summer 2018 runway shows.

detect some attributes or trends such as parts and styles Vogue’s reports matched the results of the ZARA gar-
which were mentioned in magazines. Research question ment style trends report produced by the A.I. model. It
2.2 (When using the A.I. model to analyze fashion trends is also observed that only two garment detail trends
based on the developed fashion attributes, is there any and four color trends proposed by Vogue can be
difference between the fashion trends summarized by the found from the corresponding ZARA detail and
proposed A.I. algorithm of ZARA’s products and the color trend analysis produced by the A.I. model.
forecasting analyses made by Vogue?) was also partially These results warrant further research into evaluating
answered since only two garment style trends from the trend analysis produced by the A.I. model in a
20 Textile Research Journal 0(0)

Figure 16. The result of the detected garment categories in the Spring/Summer 2018 runway shows.

more comprehensive way. The main contribution of the These factors led to the proposed A.I. algorithm failing
research compared with previous literature includes the to predict some clothing attributes correctly. Thus, if
following: our proposed A.I. model will be used by fashion pro-
Training datasets. There is no doubt that A.I. is a fessionals in analyzing fashion images, we would need
powerful approach in analyzing visual information. new datasets with correct annotations made by fashion
However, both our results and the results from previ- experts. The more accurate labeling system would help
ous research show that the training datasets used in this the A.I. model understand the concept of fashion
study can be improved in three areas: (1) The quality of attributes better. These fashion experts would also be
the image datasets needs to be improved. In our study, able to label attributes in a more comprehensive and
a large-scale fashion dataset called “DeepFashion” was less ambiguous way across different dimensions such as
adopted to teach the A.I. model the concept of fashion more specific color names and silhouette attributes.
attributes.15 However, around 60% of the fashion This would subsequently allow for the A.I. model to
attributes annotated in this “DeepFashion” dataset better align with the trend predictions of fashion edi-
contain errors. (2) Some nuanced and ambiguous tors and trend forecasting companies.
attributes in the images tend to be more difficult to More accurate A.I. model. Similar to the issue of
be categorized and recognized by the proposed A.I. training data, both our results and the results from
model. For example, the style attributes “athletic” or previous research show that A.I. models can be
“sporty” have very close meanings from a pattern rec- improved in two areas: (1) A better A.I. algorithm is
ognition perspective and presented an added challenge needed for detecting fashion attributes with higher
for the A.I. model to distinguish the difference. accuracy.26,27,35,67 (2) The current advances from
Similarly, the A.I. model does not perform well on research frontiers show that improvements of A.I.
the details and decoration on garments such as algorithms are still very promising. For instance, the
“embellished” or “embroidered.” (3) To our knowl- current advances from the computer science communi-
edge, there is no a large-scale training data of runway ty have shown that it is possible to achieve good results
images that exists currently in the research community. with few human-annotated labels by pre-training the
Similar to DeepFashion, most of the existing training A.I. models on a large unlabeled dataset and then
datasets are more focused on the daily life domain. fine-tuning the A.I. models on a smaller human-
Moreover, runway images might contain some very annotated dataset. Pre-training on large unlabeled
exaggerated attributes that are rare in daily life image datasets has the potential to improve perfor-
images. The lack of large amounts of training data in mance of recognizing images. As demonstrated by the
the runway domain might also make it difficult for the A.I. model developed by Google Research68 recently,
A.I. model to learn and predict attributes correctly. pre-training will not only teach A.I. models the concept
Shi et al. 21

Figure 17. The mapping of attribute names among our A.I. model results, Vogue Italy, Vogue France, Vogue UK, ELLE Magazine, and
Harper’s Bazaar.

Figure 18. The overlapped texture attributes among our A.I. model results, Vogue Italy, Vogue France, Vogue UK, ELLE Magazine, and
Harper’s Bazaar.

and the difference of each image, but also significantly pre-trained on a large, unlabeled dataset (developed by
reduce the requirement of annotating a large amount of Google Research) achieved 93.2% accuracy, surpassing
images for training A.I. models, which is usually expen- the human-level accuracy score of 91.2% for language-
sive and time-consuming. Furthermore, the A.I. model related tasks.69 This could be very helpful for
22 Textile Research Journal 0(0)

Figure 19. The overlapped fabric attributes among our A.I. model results, Vogue Italy, Vogue France, Vogue UK, ELLE Magazine, and
Harper’s Bazaar.

Figure 20. The overlapped shape attributes among our A.I. model results, Vogue Italy, Vogue France, Vogue UK, ELLE Magazine, and
Harper’s Bazaar.

Figure 21. The overlapped parts attributes among our A.I. model results, Vogue Italy, Vogue France, Vogue UK, ELLE Magazine, and
Harper’s Bazaar.

understanding the nuanced and ambiguous attributes proposed by researchers from the Facebook A.I.
inherent in fashion images. Research group.71 However, the proposed A.I. model
Comparison of different A.I. model architecture. We can only summarize basic apparel categories such as
did not compare the accuracy of different state-of-the- ties, bags, and umbrellas in moving model images.
art models in the current study. Only 16 images from Future works should explore how to build an A.I.
the Dolce & Gabbana collection have human- model that understands more comprehensive apparel
annotated ground truth labels, which is not valuable categories while images are moving.
for comparing the accuracy of different state-of-the- Image types. Previous work on trend analysis uti-
art models. For readers who are interested in knowing lized A.I. algorithms to extract clothing attributes
more about the performance of different state-of-the- from daily life and street images to investigate fashion
art model architectures, we refer readers to trends in clothing around the globe.12,37 Our study
Fashionpedia projects where researchers created a broke down catwalk and product images from fashion
large-scale fashion dataset fully annotated by fashion companies to deduce style trends. Both methods could
experts and then compared the accuracy of the current be useful for fashion professionals and companies.
state-of-the-art models.70 While trend predictions by fashion editors are impor-
Segmentation techniques. The results show that the tant to use as references, it has the limitation that they
quality of the segmentations generated by our A.I. only focus on a small sector of fashion that is attainable
model (Figures 7–10) is better than that of previous by few people. Keeping this in mind, it is therefore
work, as shown in Figure 1(a).13 This is mainly because important and more telling to analyze the trends
this study utilized a more advanced A.I. model employed by the wider population through analyzing
Shi et al. 23

Figure 22. The overlapped style attributes among our A.I. model results, Vogue Italy, Vogue France, Vogue UK, ELLE Magazine, and
Harper’s Bazaar.

street fashion images as well as e-commerce images on brands such as Louis Vuitton and Gucci might reveal
fast-fashion sites like this study did. We propose uni- more meaningful insights for reasons listed above.
fying and exploring both directions in the future to Modeling trends based on time series. In this study,
produce a more encompassing trend analysis. Doing only the image data of the Spring/Summer 2018 season
this would allow the A.I. model to serve as a powerful was used. However, trend forecasting is often time sen-
tool for both fast fashion and luxury companies alike sitive or time dependent. The data from a variety of
since fast-fashion companies may be more interested in seasons should be used for future studies to avoid bias.
what people are actually wearing during their daily Furthermore, sequential A.I. models such as LSTM72
lives while luxury companies may find it more helpful or BERT69 should be considered for use since Faster
to understand the major trends presented on the R-CNN is an image detection model and cannot make
runway. time-series predictions.
Comprehensive comparison. The comparison Clustering. The clustering algorithm is also experi-
between Vogue reports and ZARA product analyses mented in this study. The results of clustering are not
might be somewhat unreasonable. Vogue trend reports very promising in the study. There are two main pos-
target high-end luxury brands and products while sible reasons: (1) the pixels regions representing gar-
ZARA products are more in line with low-end fast ments are usually small, so the extracted features are
dominated by the background information; and (2) our
fashion. As mentioned previously, fast-fashion design-
image data are mainly runway show images and images
ers such as those working at ZARA may not use Vogue
from the same fashion brands share a very similar
and high-fashion magazine trends as guiding principles
background in the runway shows. However, the clus-
for their mood boards. Instead, they may focus on
tering algorithm could generate promising results if it is
what is currently popular in youth subcultures and on
used for suitable image domain or tasks, especially for
the streets of major cities. Even if they do take inspi-
diverse street style image domain, as demonstrated in
ration from the runway, their resulting designs are Bell and Bala73 and Veit et al. 74
probably very dialed down so that the average consum- Efficiency. The average time it takes for our A.I.
er feels comfortable wearing the garment since runway model to predict fashion attributes of a single image
garments are usually very over-the-top and not practi- is one second. In contrast, the average time it takes a
cal to wear. Another reason why comparing Vogue and human to visually inspect and label fashion attributes
ZARA may be unreasonable is because of their differ- of a single image is 30 seconds. The average time of a
ence in taxonomy and wording of trends. For instance, human is measured and based on three students
there may have actually been some shared attributes recruited from our fashion department. Therefore,
between the Vogue trend predictions and ZARA prod- compared to our A.I. models, it will take much
ucts that the A.I. model was not able to detect simply longer for a fashion editor or forecaster to go through
because fashion editors tend to use more refined and all the runway images in a single season.
artsy verbiage such as “baby doll” that ZARA would Cost and sustainability. Compared with hiring a
simply label as a short dress. In future studies, compar- human expert, using the proposed A.I. model in
ing the trend analyses between Vogue and luxury trend analysis is more sustainable and cost-effective.
24 Textile Research Journal 0(0)

Figure 23. (a,b) The result of the garment style and detail trends.

Fashion editors or forecasters need to manually go Unlike teaching and training humans, teaching and
through all the runway show images in every new training A.I. models to recognize fashion attributes is a
season to analyze trends, which involves a constant one-time thing. Once the model is trained, fashion com-
human expense for each season to perform this panies just need to maintain the A.I. models in a
task. This will become worse with the rapid growth production-level server, such as Amazon Web
of social networks such as Instagram, which enables Services (AWS).75 AWS is a subsidiary of Amazon pro-
current fashion trends to change faster and more viding on-demand cloud computing platforms to indi-
dynamically than ever. Many fashion bloggers are viduals, companies, and governments on a metered
creating their own fashion trends on social network pay-as-you-go basis. These cloud computing web serv-
platforms, which would be interesting to examine but ices provide a variety of basic abstract technical infra-
extremely overwhelming and difficult for editors structure and enable trained A.I. models to run 24/7.
and buyers to analyze manually given the limited The cost of maintaining the A.I. model using the AWS
time frame. service is fairly low and usually costs a few hundred or
Shi et al. 25

Figure 24. (a,b) The result of the color and fiber content trends.

thousand dollars, depending on the type of AWS ser- believed to have the ability to detect fashion attributes
vice chosen. from images with some degree of accuracy (around
75%) from static runway images with some promising
results from moving fashion models in runway videos.
Conclusion The proposed A.I. model is able to discover garment
The results from Research question 1 in this explorato- details that cannot be easily detected by the human eye
ry study indicated that the proposed A.I. model is along with the ability to run 24/7. It can also be easily
26 Textile Research Journal 0(0)

scaled up to detect fashion attributes and perform on a products may have to contain some commercial factors
large number of images – identifying not only similar- such as integrating the fashion attributes or styles of
ities but also differences among various types of gar- best-selling products from previous seasons or they
ments and connect runways to “real-ways” or vice may have to represent their company’s vision or cul-
versa. ture. Only relying on trends reports from professional
The results from Research question 2 in this explor- magazines would be too risky and cannot guarantee
atory study concluded that the proposed A.I. model is sales and revenue for the current season. In addition,
believed to have the ability to discover fashion trends fashion style preference in different locations (such as
based on the developed fashion attributes as compared Los Angeles and New York City) will impact designers’
with the trends developed in the world’s leading fash- decisions for their product lines due to the environment
ion magazines. The proposed A.I. model is also able to or culture influence. This fact further helps the under-
discover fashion trends for ZARA’s products as com- standing why fashion trends summarized from the A.I.
pared with the forecasting analyses prepared by Vogue algorithm based on ZARA’s products and the forecast-
to some degree. In addition, the proposed A.I. model ing by Vogue is not well-matched as compared to the
could detect some fashion attributes and trends that first part of the study.
fashion editors fail to discover, since the A.I. model is The study results also indicate that there is a need
able to provide a more comprehensive overview of all for further investigation of trend analysis produced by
possible fashion trends by analyzing all runway images the proposed A.I. model. For instance, the comparison
in a much more efficient way. Hence, by leveraging the between the fashion trends summarized from the A.I.
proposed A.I. models, “fashion insiders” such as fash- algorithm and the world’s leading fashion magazines
ion forecasters and designers would not have to spend (section 3.2.1) reveals that it does exist the discrepancy
hours in analyzing fashion images, producing trend between how a fashion editor names a trend and how
reports, and preparing for product development man- the A.I. model labels that same trend. Similarly, the
ually. If presented with the proposed A.I. model’s trend comparison between fashion trends summarized from
analyses, these “fashion insiders” would also be the A.I. algorithm based on ZARA’s products and the
equipped with a more comprehensive and all- forecasting in Vogue (see section The comparison
encompassing eye for trends that span across different between the fashion trends summarized from the A.I.
fashion images as opposed to only those from high-end algorithm based on ZARA’s products and the forecast-
fashion runways. Furthermore, this would allow ing by Vogue) also shows disparate information
designers to better understand and create products between Vogue and ZARA which potentially causes
that appeal to consumers, optimize their product the difference in the trends report. While the developed
assortment and ultimately generate more profits with A.I. model is not yet perfect, adoption of an A.I. model
less wasted overstock. has demonstrated promising potential to provide fash-
By further comparing the two parts of the study ion designers, trend forecasters, fashion media editors,
conducted under Research question 2, it reveals the and buyers with several benefits:
difference in interpreting trends from the proposed A.
I. model as compared with the trends from fashion 1. It can lessen the financial and environmental burden
editors and designers. The results of the first part of for professional fashion companies and
the study (trends summarized from the A.I. model vs. organizations.
the world’s leading fashion magazines) explained that 2. It can recognize garment details not easily seen by
fashion editors tend to interpret fashion trends in a the human eye.
more general way as they usually only need to summa- 3. It can help trend forecasting firms and fashion com-
rize trends based on the most frequent fashion attrib- panies work in a more efficient and holistic manner.
utes or styles in runway shows without considering 4. It can help fashion companies better understand
other resources or factors. This is one of the important their target consumers’ fashion preference by using
reasons why fashion trends summarized from the A.I. the proposed A.I. model to extract and analyze fash-
algorithm and the world’s leading fashion magazines ion attributes from best-selling product images in
are more aligned as compared with the second part of previous seasons.
the study (the trends summarized from A.I. algorithm 5. It can help fashion companies make the product
based on ZARA’s products vs. Vogue). Furthermore, (clothes) management more efficient by organizing
fashion designers need to analyze fashion trends with inventory according to the categorized fashion
deeper understanding. In a fashion company, designers attributes developed by the proposed A.I. model.
usually work with trend reports prepared by fashion
forecaster or editors combined with their own perspec- Current fashion trends are changing faster and more
tives to create commercial products. For example, their dynamically than ever. Many fashion bloggers are
Shi et al. 27

creating their own fashion trends making it very diffi- retailers (ZARA and H&M). Fashion forecasting
cult for editors and buyers to analyze them all manu- companies such as WGSN tend to produce more
ally. Additionally, thousands of style images are posted well-rounded analyses than fashion magazines.
on platforms such as Instagram every day which would They have a team dedicated to investigating a wide
be interesting to examine but extremely overwhelming variety of brands in all different tiers as well as
and difficult for fashion editors to collect and group people who travel throughout the world to observe
together, then visualize trends manually. On the other not only runway shows but also street style.
hand, fast-fashion companies such as ZARA and Therefore, comparing the proposed A.I. model’s
H&M feel the need to constantly release product lines trend analysis to that of trend forecasting companies
and shorten the typical seasonal calendar to meet con- such as WGSN will help improve the accuracy of A.
sumers’ fast-changing tastes and preferences. To do I. technology. It might also show which trend
this, they have to generate accurate trend predictions reports (if any) really influence and affect the direc-
within limited time intervals with less labor costs in tion of product lines for specific brands. Overall,
hopes of better optimizing their product offering and future studies would need to look at a much wider
reducing overstock. It is reported that H&M holds $4.3 range of organizations, people, and companies in
billion in unsold inventory.76 This not only produces a order to gain a comprehensive understanding of
major loss for the company, but also leads to a serious the industry as a whole and the complexities associ-
waste issue in which the unsold inventory ends up in a ated with it.
landfill or sometimes is even burned. The developed A. 3. Unseen trends. Due to the lack of location and geo-
I. augmented fashion trend analysis methods show they logical information associated with the images, the
can detect attributes and summarize trends automati- proposed A.I. method was not able to explore unex-
cally from a variety of resources or channels without pected or interesting trends that could exist in the
much labor costs involved. They can also be leveraged images (as mentioned in “Stage 2: Analyzing fashion
by fashion companies as a tool to improve their sus- trends”). Besides solely observing high-level fashion
tainable practices and help in generating less waste, trends, the A.I. augmented method also has the
which is an emerging issue in the industry right now. potential to capture the complex and multidimen-
sional aspects of an image that cannot be observed
by the human eye. If street style snapshots are
Limitations and future work included in future studies, it will be possible for
The key findings of this exploratory study with the the A.I. model to deduce even more complex insights
narrowly defined scope of work suggest future studies such as specific genders and age groups wearing par-
can be done to improve the accuracy of detecting fash- ticular styles/trends which would be helpful for fash-
ion attributes in a wider range of fashion images in ion companies to define their target consumers.
diverse settings. 4. Accuracy. Several inaccuracies were observed while
analyzing the Dolce & Gabbana Spring 2018 ready-
1. Street snapshots. Only runway show images and to-wear collection. Improving the accuracy of the
e-commerce photos were analyzed in this study. In proposed A.I. method is an ongoing project.
future studies, street style images should also be ana- Owing to the lack of human-annotated ground
lyzed using the proposed A.I. method. Street style is truth labels for the larger sample of runway show
less dramatic than runway style, so it may be inter- images, this study is more focused on the qualitative
esting to compare the results from both street style analysis of generated attributes. In future studies, a
and runway images. Street style images could be more quantitative analysis should be conducted on
important to explore since they are often more rep- runway show images, as demonstrated in Jia et al.70
resentative and telling than runway images, which 5. Utilization among fashion industry professionals.
can be very exaggerated and over-the-top. Another possible challenge of this study is that
2. Diversity. This study only uses the A.I. Model to designers, creative directors, etc. may not be adept,
testify one brand’s (ZARA) trend analysis compared or even willing, to adopt these new A.I. technolo-
with the predictions published in one magazine gies. It may be a severe learning curve to get this
(Vogue). More fashion brands and trend forecasting technology adopted and utilized in the fashion
companies should be included to verify the effective- industry, which is considered a very traditional
ness of the A.I. Model’s trend analysis in future industry.
studies. It would be interesting to reveal the differ-
ences between the trend reports made by fashion In summary, the research results demonstrate that
media vs. forecasting companies (Vogue and the partnership with an A.I. algorithm gives “fashion
WGSN) and the products sold by leading global insiders” a second pair of creative eyes and a more
28 Textile Research Journal 0(0)

exhaustive analysis of fashion trends. Certainly, it can 9. Bourdev L, Maji S and Malik J. Describing people: a
lessen the financial and environmental burden of poselet-based approach to attribute classification. In:
traditional labor-intensive abstraction processes IEEE international conference on computer vision, ICCV,
within fashion companies. The A.I. algorithm no Barcelona, Spain, 2011, pp. 1543–1550. Piscataway, NJ:
doubt provides a potential emerging alternative in IEEE.
10. Chen H, Gallagher A and Girod B. Describing clothing
trend forecasting in a more efficient, accurate, sustain-
by semantic attributes. In: European conference on com-
able, and cost-effective way for the fashion industry.
puter vision, 2012, pp. 609–623. Berlin, Heidelberg:
Springer.
Acknowledgement 11. Di W, Wah C, Bhardwaj A, et al. Style finder: fine-
The authors thank Serge Belongie, Claire Cardie, Kavita grained clothing style detection and retrieval. In: IEEE
Bala, and Huiju Park for their helpful feedback and discus- Conference on computer vision and pattern recognition
sion in the development of this paper. The authors thank R.J. workshops (CVPRW), 2013, pp. 8–13. Piscataway, NJ:
Lehman, Corina and Dan Lecca for providing the runway IEEE.
show images for this study. The authors thank Rohun 12. Simo-Serra E, Fidler S, Moreno-Noguer F, et al. A high
Tripathi and Omid Poursaeed for their assistance in crawling performance CRF model for clothes parsing. In: Asian
web data. conference on computer vision, 2014, pp. 64–81. Berlin,
Heidelberg: Springer.
Declaration of conflicting interests 13. Yamaguchi K, Hadi Kiapour M and Berg TL. Paper doll
The author(s) declared no potential conflicts of interest with parsing: Rrtrieving similar styles to parse clothing items.
respect to the research, authorship, and/or publication of this In: Proceedings of the IEEE international conference on
article. computer vision, 2013, pp. 3519–3526. Piscataway, NJ:
IEEE.
Funding 14. Yamaguchi K, Kiapour MH, Ortiz LE, et al. Parsing
clothing in fashion photographs. In: IEEE conference
The author(s) received no financial support or the research,
on computer vision and pattern recognition (CVPR),
authorship, and/or publication of this article.
2012, pp. 3570–3577. Piscataway, NJ: IEEE.
15. Liu Z, Luo P, Qiu S, et al. Deepfashion: Powering robust
ORCID iD
clothes recognition and retrieval with rich annotations.
Cali Chussid https://orcid.org/0000-0001-5431-7394 In: Proceedings of the IEEE conference on computer
vision and pattern recognition, 2016, pp.1096–1104.
References Piscataway, NJ: IEEE.
1. Banks L. How to Tell the Fashion Future? New York 16. Gu X, Gao F, Tan M, et al. Fashion analysis and under-
Times [internet], 26 Aug 2013, www.nytimes.com/2013/ standing with artificial intelligence. Inf Process Manag
08/27/fashion/how-to-tell-the-fashion-future.html 2020; 57(5): 102276.
2. Fiore AM and Kimle PA. Understanding aesthetics for the 17. Di W, Wah C, Bhardwaj A, et al. Style finder: fine-
merchandising and design professional. 2nd ed. New York: grained clothing style detection and retrieval. In:
Fairchild Books, 1997. Proceedings of the IEEE conference on computer vision
3. Bloomsbury Publishing. Fashion Forecasting, https:// and pattern recognition workshops, 2013, pp. 8-13.
www.bloomsbury.com/uk/fashion-forecasting- Piscataway, NJ: IEEE.
9781563678202/ (accessed 24 March 2021). 18. Corbiere C, Ben-Younes H, Rame A, et al. Leveraging
4. Ren S, He K, Girshick R, et al. Faster r-cnn: towards weakly annotated data for fashion image retrieval and
real-time object detection with region proposal networks. label prediction. In: Proceedings of the IEEE international
Adv Neural Inf Process Syst 2015; 91–99. conference on computer vision workshops, 2017, pp. 2268-
5. Bossard L, Dantone M, Leistner C, et al. Apparel classi- 2274. Piscataway, NJ: IEEE.
fication with style. In: Asian conference on computer 19. Chen H, Gallagher A and Girod B. Describing clothing
vision, 2012, pp. 321–335. Berlin, Heidelberg: Springer. by semantic attributes. In: European conference on com-
6. Pont-Tuset J. CVPR 2015: Is deep learning in fashion? puter vision, 7 October 2012, pp. 609-623. Berlin,
Scraping vision papers, jponttuset.github.io/deep-learn Heidelberg: Springer.
ing-scraping/ (accessed 10 November 2019). 20. Zhang N, Paluri M, Ranzato MA, et al. Panda: Pose
7. Hidayati SC, Hua K-L, Cheng W-H, et al. What are the aligned networks for deep attribute modeling. In:
fashion trends in New York? In: Proceedings of the 22nd Proceedings of the IEEE conference on computer vision
ACM international conference on multimedia, 2014, pp. and pattern recognition, 2014, pp. 1637-1644.
197–200. New York, USA: ACM. Piscataway, NJ: IEEE.
8. Vittayakorn S, Yamaguchi K, Berg AC, et al. Runway to 21. Yamaguchi K, Okatani T, Sudo K, et al. Mix and match:
realway: visual analysis of fashion. In: IEEE Winter joint model for clothing and attribute recognition.
Conference on applications of computer vision (WACV), Proceedings of BMVC 2015, Vol. 1, No. 2, p. 4.
5–9 Jan 2015, pp. 951–958. Piscataway, NJ: IEEE. Swansea, UK: BMVA Press,
Shi et al. 29

22. Park S, Nie BX and Zhu SC. Attribute and-or grammar and pattern recognition, 2017, pp. 7310–7311. Piscataway,
for joint parsing of human pose, parts and attributes. NJ: IEEE.
IEEE Trans Pattern Anal Mach Intell 2017; 40(7): 37. Matzen K, Bala K and Snavely N. Streetstyle: exploring
1555–1569. world-wide clothing styles from millions of photos.
23. Han X, Wu Z, Huang PX, et al. Automatic spatially- ArXiv:1706.01869. Epub ahead of print 2017.
aware fashion concept discovery. In: IEEE international 38. Cheng Z-Q, Wu X, Liu Y, et al. Video2Shop: exact
conference on computer vision (ICCV), 2017, pp. 1472– matching clothes in videos to online shopping images.
1480. Piscataway, NJ: IEEE. In: 2017 IEEE conference on computer vision and pattern
24. Singh KK and Lee YJ. End-to-end localization recognition (CVPR), 2017, pp. 4169–4177. Piscataway,
and ranking for relative attributes. Berlin, Heidelberg: NJ: IEEE.
ECCV 2016; 6: 753–769. 39. Al-Halah Z, Stiefelhagen R and Grauman K. Fashion
25. Jia M, Zhou Y, Shi M, et al. A deep-learning-based fash- forward: forecasting visual style in fashion. In:
ion attributes detection model. arXiv preprint Proceedings of the IEEE international conference on com-
arXiv:1810.10148. 24 Oct 2018. puter vision, 2017, pp. 388–397. Piscataway, NJ: IEEE.
26. Chen Q, Huang J, Feris R, et al. Deep domain adaptation 40. Song Z, Wang M, Hua X-S, et al. Predicting occupation
for describing people based on fine-grained clothing via human clothing and contexts. ICCV 2011; 1084–1091.
attributes. In: IEEE conference on computer vision and 41. Kwak IS, Murillo AC, Belhumeur PN, et al. From bikers
pattern recognition (CVPR), 2015, pp. 5315–5324. to surfers: visual recognition of urban tribes. In: British
Piscataway, NJ: IEEE. machine vision conference, September 2013, Vol. 1, p. 2.
27. Dong Q, Gong S and Zhu X. Multi-task curriculum Swansea, UK: BMVA Press.
transfer deep learning of clothing attributes. In: IEEE 42. Chen K-T and Luo J. When fashion meets big data: dis-
winter conference on applications of computer vision criminative mining of best-selling clothing features.
(WACV), 2017, pp. 520–529. Piscataway, NJ: IEEE. WWW (companion volume) 2017; 15–22.
28. Hadi Kiapour M, Han X, Lazebnik S, et al. Where to 43. He R and McAuley J. Ups and downs: modeling the
buy it: matching street clothing photos in online shops. visual evolution of fashion trends with one-class collab-
Proceedings of the IEEE international conference on com- orative filtering. Proceedings of the 25th international con-
puter vision, 2015, pp. 3343–3351. Piscataway, NJ: IEEE. ference on world wide web, 2016, pp. 507–517.
29. Ma Y, Jia J, Zhou S, et al. Towards better understanding International World Wide Web Conferences Steering
the clothing fashion styles: a multimodal deep learning Committee.
approach. Palo Alto, USA: AAAI 2017; 38–44. 44. Song Z, Wang M, Hua X-S, et al. Predicting occupation
30. Simo-Serra E and Ishikawa H. Fashion style in 128 via human clothing and contexts. ICCV 2011; 1084–1091.
floats: joint ranking and classification using weak data 45. Rostamzadeh N, Hosseini S, Boquet T, et al. Fashion-
for feature extraction. In: Proceedings of the IEEE con- gen: the generative fashion dataset and challenge. arXiv
ference on computer vision and pattern recognition, 2016, preprint arXiv:1806.08317. 21 Jun 2018.
pp. 298-307. Piscataway, NJ: IEEE. 46. Zou X, Kong X, Wong W, et al. Fashionai: a hierarchical
31. Jiang S, Shao M, Jia C, et al. Consensus style centralizing dataset for fashion understanding. In: Proceedings of the
auto-encoder for weak style classification. In: IEEE conference on computer vision and pattern recogni-
Proceedings of the AAAI conference on artificial intelli- tion workshops, 2019. Piscataway, NJ: IEEE
gence, 21 Feb 2016, Vol. 30, No. 1. Palo Alto, USA: 47. Ge Y, Zhang R, Wang X, et al. Deepfashion2: a versatile
AAAI. benchmark for detection, pose estimation, segmentation
32. Hsiao WL and Grauman K. Creating capsule wardrobes and re-identification of clothing images. In: Proceedings
from fashion images. In: Proceedings of the IEEE confer- of the IEEE conference on computer vision and pattern
ence on computer vision and pattern recognition, 2018, pp. recognition, 2019, pp. 5337-5345. Piscataway, NJ: IEEE.
7161-7170. Piscataway, NJ: IEEE. 48. Zheng S, Yang F, Kiapour MH, et al. Modanet: a large-
33. Lee H, Seol J and Lee SG. Style2vec: representation scale street fashion dataset with polygon annotations. In:
learning for fashion items from style sets. arXiv preprint Proceedings of the 26th ACM international conference on
arXiv:1708.04014. 14 Aug 2017. multimedia, 15 Oct 2018, pp. 1670–1678. New York,
34. Vaccaro K, Shivakumar S, Ding Z, et al. The elements of USA: ACM.
fashion style. In: Proceedings of the 29th annual sympo- 49. Ji X, Wang W, Zhang M, et al. Cross-domain image
sium on user interface software and technology, 16 Oct retrieval with attention modeling. ArXiv:1709.01784.
2016, pp. 777–785. Epub ahead of print 2017.
35. Huang J, Feris RS, Chen Q et al. Cross-domain image 50. Kuo Y-H and Hsu WH. Feature learning with rank-
retrieval with a dual attribute-aware ranking network. based candidate selection for product search.
Proceedings of the IEEE international conference on com- Proceedings of the IEEE conference on computer vision
puter vision, 2015, pp. 1062–1070. Piscataway, NJ: IEEE. and pattern recognition, 2017, pp. 298–307. Piscataway,
36. Huang J, Rathod V, Sun C, et al. Speed/accuracy trade- NJ: IEEE.
offs for modern convolutional object detectors. In: 51. Wang X, Sun Z, Zhang W, et al. Matching user photos to
Proceedings of the IEEE conference on computer vision online products with robust deep features. Proceedings of
30 Textile Research Journal 0(0)

the 2016 ACM on international conference on multimedia 63. Harper’s Bazaar. The top spring 2018 runway trends to
retrieval, 2016, pp. 7–14. New York, USA: ACM. know. Harper’s Bazaar [magazine on the internet], www.
52. Vogue. Summer 2018 fashion trends: Summer dresses’ harpersbazaar.com/fashion/trends/g12176306/spring-
styles and patterns. Vogue [magazine on the internet]. 2018-fashion-trends/
13 April 2018, www.vogue.it/en/fashion/trends/2018/04/ 64. Python. Available from: https://www.python.org/
13/summer-2018-fashion-trends-summer-dresses-styles- 65. Pandas. Available from: https://pandas.pydata.org/
patterns/ 66. Matplotlib. Available from: https://matplotlib.org/
53. Vogue. The ten color trends for Spring Summer 2018. 67. Du X, Lin TY, Jin P, et al. Spinenet: learning scale-
Vogue [magazine on the internet]. 12 January 2018, permuted backbone for recognition and localization. In:
www.vogue.it/en/fashion/trends/2018/01/12/the-ten- Proceedings of the IEEE/CVF conference on computer
color-trends-for-spring-summer-2018/ vision and pattern recognition, 2020, pp. 11592–11601.
54. Vogue. 10 Top Trends for Spring Summer 2018. Vogue Piscataway, NJ: IEEE.
[magazine on the internet], www.vogue.it/en/fashion/ 68. Google AI Blog. Advancing self-supervised and semi-
trends/2017/10/04/best-trend-fashion-spring-summer- supervised learning with SimCLR, https://ai.googleblog.
2018/?refresh_ce= com/2020/04/advancing-self-supervised-and-semi.html
55. ELLE. ELLE.Com’s Comprehensive Guide To The (accessed 8 April 2020).
69. Google AI Blog. Open Sourcing BERT: state-of-the-art
Biggest Trends Of Spring 2018. Elle [magazine on the
pre-training for natural language processing, http://ai.
internet], www.elle.com/fashion/trend-reports/
googleblog.com/2018/11/open-sourcing-bert-state-of-art-
g12205305/spring-fashion-trends-2018/
pre.html (accessed 20 April 2019).
56. He K, Zhang X, Ren S, et al. Deep residual learning for
70. Jia M, Shi M, Sirotenko M, et al. Fashionpedia: ontolo-
image recognition. In: Proceedings of the IEEE conference
gy, segmentation, and an attribute localization dataset.
on computer vision and pattern recognition, 2016, pp. 770-
In: European conference on computer vision, 23 Aug, 2020,
778. Piscataway, NJ: IEEE. pp. 316-332. Cham: Springer.
57. Tensorflow. Available from: www.tensorflow.org/ 71. He K, Gkioxari G, Dollár P, et al. Mask R-CNN, https://
58. Supplementary information of dataset modification and arxiv.org/abs/1703.06870v3 (2017).
model structure, https://textile-research-journal.s3.amazo 72. Hochreiter S and Schmidhuber J. Long short-term
naws.com/Supplementary+Information.pdf memory. Neural Comput 1997; 9(8): 1735–1780.
59. Lin T-Y, Maire M, Belongie S, et al. Microsoft coco: 73. Bell S and Bala K. Learning visual similarity for product
common objects in context. In: European conference on design with convolutional neural networks. New York,
computer vision, 2014, pp. 740–755. Berlin, Heidelberg: USA: ACM Trans Graphics 2015; 34(4): 1-0.
Springer. 74. Veit A, Belongie S and Karaletsos T. Conditional simi-
60. Delectron. Available from: https://github.com/facebook larity networks. In: Proceedings of the IEEE conference
research/detectron on computer vision and pattern recognition, 2017, pp. 830-
61. Vogue. 20 trends for Spring/Summer 2018. Vogue [mag- 838. Piscataway, NJ: IEEE.
azine on the internet], www.vogue.fr/fashion/fashion- 75. Amazon Web Services. Available from: https://aws.
inspiration/diaporama/fwss18-spring-summer-2018-fash amazon.com/
ion-trends/46478 76. Paton E. H&M, a fashion giant, has a problem: $4.3
62. Vogue. The Top Trends of Spring/Summer 2018. Vogue billion in unsold clothes. The New York Times [newspa-
[magazine on the internet], www.vogue.co.uk/gallery/ per on the internet], 30 March 2018, www.nytimes.com/
spring-summer-2018-trends 2018/03/27/business/hm-clothes-stock-sales.html