key: cord-0465216-lvqnpisc authors: Zou, Xingxing; Wong, Waikeung title: fAshIon after fashion: A Report of AI in Fashion date: 2021-05-07 journal: nan DOI: nan sha: bcf6f6e1d959a713347c8114f671cd0f46205a1e doc_id: 465216 cord_uid: lvqnpisc In this independent report fAshIon after fashion, we examine the development of fAshIon (artificial intelligence (AI) in fashion) and explore its potentiality to become a major disruptor of the fashion industry in the near future. To do this, we investigate AI technologies used in the fashion industry through several lenses. We summarise fAshIon studies conducted over the past decade and categorise them into seven groups: Overview, Evaluation, Basic Tech, Selling, Styling, Design, and Buying. The datasets mentioned in fAshIon research have been consolidated on one GitHub page for ease of use. We analyse the authors' backgrounds and the geographic regions treated in these studies to determine the landscape of fAshIon research. The results of our analysis are presented with an aim to provide researchers with a holistic view of research in fAshIon. As part of our primary research, we also review a wide range of cases of applied fAshIon in the fashion industry and analyse their impact on the industry, markets and individuals. We also identify the challenges presented by fAshIon and suggest that these may form the basis for future research. We finally exhibit that many potential opportunities exist for the use of AI in fashion which can transform the fashion industry embedded with AI technologies and boost profits. In this independent report -fAshIon after fashion, we examine the development of fAshIon (artificial intelligence (AI) in fashion) and explore its potentiality to become a major disruptor of the fashion industry in the near future. To do this, we investigate AI technologies used in the fashion industry -through several lenses. We summarise fAshIon studies conducted over the past decade and categorise them into seven groups: Overview, Evaluation, Basic Tech, Selling, Styling, Design, and Buying. The datasets mentioned in fAshIon research have been consolidated on one GitHub page for ease of use 1 . We analyse the authors' backgrounds and the geographic regions treated in these studies to determine the landscape of fAshIon research. The results of our analysis are presented with an aim to provide researchers with a holistic view of research in fAshIon. As part of our primary research, we also review a wide range of cases of applied fAshIon in the fashion industry and analyse their impact on the industry, markets and individuals. We also identify the challenges presented by fAshIon and suggest that these may form the basis for future research. We finally exhibit that many potential opportunities exist for the use of AI in fashion which can transform the fashion industry embedded with AI technologies and boost profits. According to AI in Fashion Market Research Report 2021 [112] , under the cumulative impact of COVID-19, global spending on AI in the fashion market is expected to grow from USD 229 million in 2019 to USD 1,260 million by 2024, at a Compound Annual Growth Rate (CAGR) of 40.8% during the forecast period. Global expenditure on AI in the fashion market is expected to grow from USD 352.58 million in 2020 to USD 825.19 million by the end of 2025. In this report, we analyse the fAshIon research and their applications, explore the potential of fAshIon to transform the fashion industry and discuss the challenges and opportunities. Firstly, we investigate fAshIon research papers reported over the past decade. We search for papers focused on the applications of AI technologies in fashion involving the databases such as IEEE Xplore Digital Library 2 , Google Scholar 3 , and Arxiv 4 . The earliest paper reviewed in this report was published on 31 March 2008 [42] ; the most recent paper was published on 9 March 2021. Five hundred twenty-one papers are analysed, covering 1,465 authors affiliated with 383 different entities (different departments of the same university or company are not regarded as different affiliations; for example, Amazon Lab126, Amazon Visual Search and AR Group are all counted as Amazon). As shown in Figure 1 , China, the United States of America, Singapore, India, Japan, and Germany are identified as the hottest regions of fAshIon research. We present region information here to emphasise that fashion is extremely complicated which may be treated in different ways in different regions. Taking an example of the attribute datasets collected for fashion recognition, the data distribution is greatly different. 'Paisley', a specific type of 'print', which is a general attribute in the fashion datasets collected from India but may be a long-tailed attribute in the datasets collected from other regions. In the following, more detailed region information will be presented. This region information provides the interested parties in fAshion with the hidden clues of the application popularity in different regions and the potential opportunities. Unlike the general 'Review Paper', the methods and techniques of related fashion tasks using AI have not been presented in detail one by one. We organise fAshIon studies according to their applications in the fashion industry to clearly show the gap between existing research and practical requirements. Meanwhile, we point out the hottest topics according to the word cloud based on the titles of papers. Statistical results of these papers, including the authors, the author's affiliations and their corresponding regions, the years of publication, most popular papers, most active researchers, etc., are presented to provide rich information of fAshIon for researchers. By presenting the analysis of fAshIon research, we hope to provide the information about 1. the gap between existing research and practical requirements; and 2. the hottest topics of each practical direction. As a result, researchers who have interests in this filed can quick start their fAshIon research by paying attention on the most active authors, reading the most porpular papers, and utlizing the summary of published fashion datasets 5 . Details can be found in Section 1. Additionally, we also investigate related products and services provided by 126 companies and start-ups. A wide range of cases of applied fAshIon in the fashion industry are reviewed and analysed in terms of their impact on the industry, markets, and individuals. By presenting the analysis of fAshIon applications, we hope to present the following key information: 1. existence of current types of fAshIon applications; 2. many technologies not being converted to practical application; 3. existence of barriers between technical frameworks and business models. Details can be found in Section 2. Finally, we draw conclusions about the challenges and the opportunities that exist in fAshIon, which may form the basis for future research. This detailed analysis is described in Section 3 and in Section 4. To our knowledge, 521 papers related to fAshIon have been published in the last 13 years. Figure 2 indicates that the overall trend of fAshIon-related papers is one of growth. The first identified paper on fAshIon paper was published in 2008. Although there is no record of fAshIon papers published in 2009 or 2010, the number of papers has been increasing continuously since 2011. The year 2017 included a particularly conspicuous step in the growth of fAshIon research, which is consistent with the conclusion of the fashion market in [112] . During the years 2017, 2018, and 2019, 80, 130, and 135 papers on this topic were published, respectively. The number of papers published in 2020 is lower compared with those in 2019 and 2018 and almost equal to the number published in 2017. One possible reason for this is the influence of the COVID-19 pandemic, which swept the world in 2020. Additionally, the gap between these research results and the market's expectations also may have directly dampened the researchers' enthusiasm to a certain extent. After carefully reading about these 521 papers, as shown in Figure 3 , we categorised them into 7 groups, namely: Overview, Evaluation, Basic Tech, Selling, Styling, Design, and Buying. This taxonomy was inspired by [187] , which divides fAshIon research into groups based on existing roles in the fashion industry, namely the fashion lover, the stylist, and the designer. We organise these papers in this taxonomy to better illustrate the relationship between computer vision problems and fashion tasks. As shown in Figure 3 , fashion tasks are extremely complicated since they would be affected by many factors. For example, 'compatibility learning' [44] which may help to reduce the workload of a stylist by generating the outfit composition automatically. Fitting a model to an outfit dataset is not difficult. The challenge is how to explain the generated results. Fashion is subjective. In most situations, a stylist needs a reasonable explanation to persuade customers to believe something or change their mind when they hold a different opinion. If no explanation to underpin the decision such as evaluation or recommendation, the decision is not convincing enough for the customers to accept. Naturally, Figure 4 : Distribution of published fAshIon papers by 7 groups, namely, Overview, Evaluation, Basic Tech, Selling, Styling, Design, and Buying. when tackling the compatibility learning task, how to give a concrete reason for the results needs to be taken into consideration. All in all, as the papers are categorised in this way, it is easier for identifying corresponding roles in the fashion industry and convenient for comparing existing research with practical requirements. Based on this taxonomy, we further analyse the 521 papers. The distribution of published fAshIon papers by these 7 groups is shown in Figure 4 . It can be found that most papers were focused on solving problems in the group of 'Selling'. Based on the research interest, these groups sorted from high to low are 'Selling', 'Design', 'Styling', 'Basic Tech', 'Buying', 'Overview', and 'Evaluation'. Furthermore, as shown on the right side of Figure 4 , we can find more interesting information of fAshIon research. Papers in the group of 'Overview' were relatively evenly distributed by year. Very few researchers have focused on defining the evaluation standard for fashion tasks. Although solving the problems in the group of 'Basic Tech' is not so popular comparing with the other groups fAshIon research, it has still received attention every year. Additionally, we analyse the group of 'Selling', 'Styling', 'Design', and 'Buying' together since all these four groups are closely related to corresponding roles in the fashion industry. Researchers focused on solving the problems in 'Selling' at the earliest (in 2011) and gradually turned their attention to the research on 'Styling' (in 2012), 'Design' (in 2017), and 'Buying' (in 2017). 'Selling' is the most popular topic among these four groups before the year 2018. The sum of papers in these three groups, i.e. 'Styling', 'Design', and 'Buying', is 27 (i.e., 13 + 9 + 5 = 27) in 2017. Meanwhile, 53 papers in total related to the tasks in the group of 'Selling' were published in 2017. In other words, in 2017, papers published in the 'Selling' group was over the total number of the 'Styling', 'Design', and 'Buying' group, i.e., 43 versus 27. In the following year 2018, the compared numbers are changed to 53 versus 62 (i.e. 25 + 35 + 2 = 62). Even though the number of papers under 'Selling' group still accounted for the largest proportion of published papers among these four groups in 2018, the situation has started to change in 2019. The numbers of published papers in the 'Selling', 'Styling', and 'Design' group were very close and almost equal in 2019, which is 38, 38, and 39, respectively. In other words, researchers turned their attention from solving problems in the 'Selling' group into tackling tasks related to 'Styling' and 'Design'. 'Design' became the most popular topics among these four groups after 2019. In addition, it is worth mentioning that, although research in 'Buying' is relatively limited, it is important to the fashion industry as the same as 'Selling', 'Styling', and 'Design'. On the basis of the analysis above, we conclude that 1. the overall trend of the fAshIon papers is growing; 2. research on proposing the evaluation protocol for specific fashion tasks has not yet raised attention; 3. research on solving problems in the group of 'Selling' was popular before and slowly lost attention; 4. research on tackling tasks in the group of 'Styling' and 'Design' becomes increasingly hot in recent years; 5. research on problems related to 'Buying' was overall increased year by year. Papers in the group of 'Buying' are still relatively few which is not in directly proportional to its importance to the fashion industry. In the following sections, we present our detailed analysis according to each defined group. The definition of each group is given first, followed by a word cloud based on the titles of the papers in the group as appropriate. Then, we present the statistical analysis of papers organised by year of publication, authorial affiliations, the corresponding regions, and the authors. Although we do not introduce the methods proposed in these papers, the most active authors and the most popular papers (according to the number of citations) are listed as a reference for researchers. Figure 5 : Distribution of Overview papers. These include 4 papers from China, 2 from India, 1 from Ireland, and 1 from Singapore. The number reflects the region of the first author's affiliation rather than the nationality of the first author. The papers in this group are review papers [33, 65, 98, 11, 134] that summarise the developments in fAshIon technology in different areas. Overview papers can help researchers to quickly understand the current situation of fAshIon research and familiarise them with the corresponding methods, related datasets, baseline approaches, and evaluation protocols. For example, [65] presents a summary of the various retrieval techniques used in clothing retrieval, and [11] provides an analysis of the four main tasks in fashion: fashion detection, including landmark detection, fashion parsing, and item retrieval; fashion analysis, consisting of attribute recognition, style learning, and popularity prediction; fashion synthesis, which involves style transfer, pose transformation, and physical simulation; and fashion recommendation, which comprises fashion compatibility, outfit matching, and hairstyle suggestion. [11] also presents the benchmark datasets and the evaluation protocols of each task. [33] summarises the use of a recommendation system in situations using different methods such as End-to-End, Implicit Feedback-based, Weak Appearance Feature-based, Semantic Attribute Region-guided, and Using Adversarial Feature Transformer; then, the paper presents some approaches related to understanding aesthetics and realising personalisation in fAshIon. The ratio of Overview papers to the total number of published fAshIon papers is around 1.5%. As shown in Figure 5 , 4 papers from China, 2 from India, 1 from Ireland, and 1 from Singapore are identified. These detailed numbers are derived from the regions of the first author's affiliation rather than the nationality of the first author; for example, if the first author is affiliated with Amazon Lab126, the paper is counted as being from the United States. Although not takes a large proportion of the fAshIon research, it still plays a role for researchers to quickly have a holistic view of fAshIon from the pespective of technology. Evaluation protocols are important for research. Unlike the studies of art, which are extremely subjective, science is quantifiable. From a research perspective, science undoubtedly needs objective indicators to demonstrate the advantage of one approach to accomplish a specific task. It is unfair to evaluate a paper as better than another based on subjective criteria. For example, 'accuracy' is one of the criteria used to evaluate whether a recognition model is good. Demonstrating the advantage of one model based solely on the retrieved results is not possible. However, there is a huge contradiction here. As has been mentioned, fashion is closer to art than science. The consideration of how to design an evaluation protocol for fAshIon research with a high degree of subjectivity has become challenging. Very few research have specifically investigated the evaluation protocols in fAshIon research. The most mainstream evaluation indicators have been adopted directly from corresponding computer science tasks, such as accuracy for fashion recognition (image recognition tasks) or inception score [7] for fashion generation (image generation tasks). To our knowledge, [128] is the only paper that describes the process of conducting research from the perspective of evaluation protocol. (Note: the use of 'only' does not mean that no other fAshIon evaluation indicators have been introduced in other papers. Fill-in-the-blank accuracy and compatibility prediction AUC [152, 14, 156] are two evaluation indicators that have been introduced for the assessment of fashion compatibility. The papers in which they are introduced have adopted unique fAshIon indicators, but this is not the main Figure 6 : Word cloud based on the titles of papers in the Basic Tech group. Words such as 'the', 'a', and 'toward' were manually deleted. Some words with similar meanings, e.g., 'clothing' and 'clothings' or 'model' and 'models', were grouped together. Figure 7 : Distribution of Basic Tech papers. These include 34 papers from China, 13 from the USA, 6 from Singapore, 3 from Japan, 2 from Canada, 2 from India, 2 from Germany, 2 from the UK, 1 from Russia, 1 from Switzerland, 1 from Colombia, 1 from Korea, and 1 from Australia. goal of the research.) [128] , published in 2019 by True Fit 6 in the USA, is mainly focused on solving the problem of evaluating a fashion recommendation system by including multiple metrics that are relevant to fashion and performing within segments of users with different interaction histories. The third fAshIon research group is Basic Tech. Papers in this group are mainly focused on the basic technology used to process fashion images [149, 89, 186, 6, 75, 32, 60, 173, 81, 87, 77, 159, 147, 100, 166, 85, 158, 104, 26, 99, 164, 97, 163, 162, 155] . It is also a foundation of all computer vision tasks. To provide a snapshot of the research in this group, we present a word cloud based on these research titles. Words such as 'the', 'a', and 'toward' were manually deleted, while some words with similar meanings, e.g., 'clothing' and 'clothings' or 'model' and 'models', were grouped together. It can be seen in Figure 6 that the keywords in this group are Parsing, Fashion, Clothing, Human, Segmentation, and Landmark. We conclude that the research in Basic Tech focus on image-level processing, e.g., clothing parsing, landmark detection, key point, and apparel detection. The ratio of Basic Tech papers to all fAshIon papers is around 13.2%. (12), Xiaodan Liang (11) , Xiaohui Shen (8), Si Liu (7), Jiashi Feng (6), Ping Luo (5), Jianchao Yang (5), Ke Gong (4), Jian Dong (4), Kota Yamaguchi (4), and M. Hadi Kiapour (4). These works form the foundation of fAshIon research, especially in the application of functions to complicated images such as crowd images or images from social media. These first three groups, Overview, Evaluation, and Basic Tech, are more general and not so closely related to roles in the fashion industry. Next, we will introduce the tasks involved with similar duties of certain roles in fashion. In this section, we introduce the tasks involved with similar duties of certain areas in fashion, namely Selling, Styling, Design, and Buying. The seller takes a main role in the fashion retail sector to sell products. A good seller in fashion needs to recommend accurate products to his or her customers. Their aim is to increase sales as much as possible. The basic requirements for the seller are therefore: • That he or she can recognise the detailed attributes of fashion products and know how to find similar items according to a general description using simple words provided by his or her customers. • That he or she can accurately recommend fashion products according to observations on both his or her customers and of trends. • That he or she can reply quickly to his or her customers about whether a particular item or size is available or out of stock. : Distribution of Selling papers. This group includes 53 papers from China, 41 from the USA, 21 from India, 16 from Singapore, 10 from Japan, 9 from Korea, 9 from Germany, 9 from the UK, 3 from Poland, 3 from Canada, 3 from the Netherlands, 2 from France, 2 from the Philippines, 2 from Portugal, 2 from Sweden, 2 from Brazil, 2 from Vietnam, 2 from Belgium, 1 from Malaysia, 1 from Turkey, 1 from Spain, 1 from Austria, 1 from Switzerland, 1 from Australia, and 1 from Italy. According to these requirements, we identify many research papers that aim to provide methods to create online virtual sellers. These papers are then categorised into the Selling group [4, 18, 115, 59, 170, 105, 68, 131, 41, 56, 161, 88, 165, 154, 103, 92, 113, 143, 43, 58, 74, 35, 12, 36, 111, 40, 71, 188, 39, 9] . It can be seen in Figure 8 that the keywords of this group are Fashion, Clothing, System, Image, Attributes, Retrieval, Recommendation, and Classification. We conclude that the research in Selling has been focused on marketing fashion products like a seller, e.g., by using fashion item recognition, recommendation, and retrieval. The ratio of Selling papers to the total number of papers is around 37.8 %; accounting for the largest proportion of research among all of the seven groups identified in the current papers. This group includes 197 papers. As shown in Figure 9 , these include 53 papers from China, 41 from the USA, 21 from India, 16 from Singapore, 10 from Japan, 9 from Korea, 9 from Germany, and 9 from the UK. China, the USA, and India are therefore identified as the hottest regions for this type of research. Large bodies of research were published in 2017, 2018, and 2019. Attribute recognition and image retrieval are identified as the hottest topics in this group. (11) This body of work aims to provide a better shopping experience to online shoppers, thus improving the performance of sales. In addition, this kind of research provides a foundation for higher-level tasks. The ability to execute fine-grained fashion attributes recognition comprises the basic knowledge needed to deal with higher-level tasks such as mixing and matching. Generally speaking, a stylist who provides styling advice to customers should have good beauty sense and an ability to provide personal styling services to individual customers. They should be able to recognise fashion attributes. The basic requirements for a stylist are: • That he or she has a good sense of clothing aesthetics, such as a sense of colour, a sense of texture, and a sense of silhouette, and can easily create a well-composed outfit. (Understand Clothing Aesthetic) • That he or she knows how to make an outfit attain visual balance for a given customer and mix and match according to different situations, such as for an occasion, the seasons, or body figure. from Singapore, 6 from India, 5 from Japan, 3 from the UK, 3 from Germany, 2 from Belgium, 1 from Indonesia, 1 from Romania, 1 from Greece, 1 from France, 1 from Egypt, 1 from the Netherlands, and 1 from Canada. • That he or she always stays aware of fashion trends and can apply these trends to his or her work. According to these requirements, we identify many research papers that aim to provide online styling services. We then categorise these papers into the Styling group [49, 153, 52, 2, 121, 101, 180, 69, 122, 90, 79, 183, 146, 133, 73, 139, 76, 64, 16, 3, 90, 119, 141, 91, 156, 138, 10, 179, 44, 189] . It can be seen in Figure 10 that the keywords of this group are Fashion, Clothing, System, Image, Attributes, Retrieval, Recommendation, and Classification. We conclude that the research in Styling has been focused on providing online styling services, e.g., by using fashion compatibility learning, outfit creation, and recommendation. The ratio of Styling papers to the total number of papers is around 20.7%. As shown in Figure 11 , this group comprises 108 papers, including 37 papers from China, 36 from the USA, 9 from Singapore, 6 from India, and 5 from Japan. China, the USA, and Singapore are identified as the hottest regions for this type of research. Large bodies of research were published in 2017, 2018, and 2019. Compatibility learning and outfit recommendation are identified as the hottest topics in this group. (7), Tat-Seng Chua (5), Xun Yang (4), Yunshan Ma (4), Xuemeng Song (4), Xingnan He (4), Reza Shirvany (4), Urs Bergmann (4), and Jun Ma (4) . Works in this group focus on duties, such as those of a stylist in the fashion industry. The works aim to provide personalised online styling services that can cross-sell to increase the exposure rate of products (outfit recommendation) and to create new business models that offer either an online mix-and-match service or complete outfit boxes based on the personal preference of the customer (do personal styling). In the fashion industry, a good design depends on the abilities of a fashion designer. Generally speaking, the designer is most closely related to the stylist and needs to have similar aesthetic abilities to those of the stylist. More importantly, a designer should know not only what beauty is but also how to create it. The unique requirements for a designer are: • That he or she has the ability to transform highly abstract concepts or stories into fashion items (e.g., garments, accessories, bags, shoes, etc.) through the language of clothing. • That he or she has the ability to transfer images or feelings from his or her brain to papers via his or her hand drawing or computer software. • That he or she has the ability to create new designs based on a theme or conditioned by different constraints, such as the seasons, the trends, or the DNA of the brand (Brand DNA is the essence of your identity as a business). • That he or she always remains aware of fashion trends and can apply these trends to his or her work. from Germany, 5 from the UK, 5 from Japan, 5 from Switzerland, 3 from France, 3 from India, 2 from Singapore, 2 from Turkey, 2 from Korea, 2 from Russia, 2 from Spain, 1 from the Netherlands, 1 from Poland, 1 from Italy, and 1 from Sweden. According to these requirements, we identify many research papers that aim to help fashion designers, which we categorise into the Design group [ Figure 12 that the keywords of this group are Fashion, GAN, Cloths, Design, Image, Generative, Adversarial, Try-On, and Synthesis. We conclude that the research in Design focus on fashion generation, virtual try-on, clothing synthesis. The ratio of Design papers to the total number of papers is around 21.3%. As shown in Figure 13 In Table 4 , aside from the number of papers published by these top authors in the group, we also present the top 3 related papers according to the number of citations. Particularly, we emphasise a paper that has received much attention: 'Disentangled Person Image Generation [107] (7) Most of current works focus on helping designers or online retailers to visually demonstrate their ideas, e.g., attribute editing [23, 140, 54] , 3D garment generation [118, 5, 169, 150, 37, 38] , virtual try-on [125, 21, 72, 20, 13, 45] , etc. These methods can speed up the process of design in terms of time saving on design detail modification to a certain extent. Furthermore, material waste can be reduced since making up a physical sample for each design is not necessary. In addition, marketing materials such as videos, brochures, flyers, etc. for promoting fashion products of retailers can be generated by using GAN models and thus marketing costs involved in shooting, promotional material design, etc. can be reduced. Certainly, with a holistic view of the 'Design' research, there are many interesting but challenging tasks remain. Can AI models create new designs based on a theme or subject to different predetermined constraints, such as seasons, trends, brand images, etc.? Can AI models transform highly abstract concepts or themes into fashion items, such as garments, accessories, bags, shoes, etc.? The last group is Buying. In the fashion industry, a buyer performs the action of buying the right products for fashion brands and retailers in the retail sector. Unlike a stylist or a designer, a buyer's work depends on making the decision to buy the right products in the right quantities. The unique requirements for a buyer are: • That he or she can draw conclusions from the records of previous sales. • That he or she has great insight into the market and the ability to make predictions. • That he or she is very experienced in fashion retail and fully knows the market trends. According to these requirements, we identify many research papers on the subject of Buying [53, 110, 1, 144, 124, 109, 151, 142, 132] . It can be seen in Figure 14 that the keywords of this group are Fashion, Trends, Discovering, Visualisation, Prediction, and Social Media. We conclude that the research in Buying focus on trend prediction and culture discovering. The ratio of Design papers to the total number of papers is around 5.2%. As shown in Figure 15 In addition to our research on fAshIon, we also focus on related companies and start-ups. Based on our investigation, 126 companies and start-ups that offer fAshIon-related services, applications, or technology support have emerged in recent years. For example, Echo Look by Amazon ( Figure 16 ) provides customers with a special online service that compares two uploaded outfits and selects the better outfit using AI. Thus, the user can determine the best look for the day. The technology used in this application belongs to the previously described Styling group. Other products and functions, such as Fashion Box provided by STITCH FIX 23 , utilise similar technologies. After completing a style quiz, the customer receives a box of clothing items recommended for them, providing a different shopping experience from before. Some companies, such as Alibaba, focus on outfit generation and recommendation, presenting outfit compositions on the checkout page for cross-selling. In this section, we introduce the opportunities for fAshIon based on an analysis of the products or services provided by the investigated companies. Existing applications or products in the industry are presented briefly. Detailed information can be found on their official websites. We do not present the statistical results for these investigated companies because these numbers are largely inconsequential and because the website links of some of the investigated companies are now invalid, such as that of Glitch 24 , an AI clothing brand founded by computer scientists turned fashion designers at MIT. We review a wide range of cases of applied fAshIon in the fashion industry and categorized them according to their corresponding technologies (i.e., Selling, Styling, Design, Buying). We summarise the current situations in fAshIon and analyse its impact on the industry, markets and individuals in the end of this section. Fashion Searching and Complete the Look -'Selling': Technology can help customers to search for fashion items immediately conditioned on uploaded images. Utilising the recognition model to perform attributes tagging can also speed up the process of lunching new products. Additionally, algorithms can make fashionable recommendations for achieving a complete look, a feature that may be widely applicable in cross-selling. A case study of Fashion Searching and Complete the Look by MACTY (Figure 17 ) reveals that the platform can enable one-click searching to obtain a full outfit. This technology can help to inspire users who are considering what to wear and how to combine pieces by automatically creating the perfect look. More information can be found at catchoom 25 Stylist Service -'Styling': As shown in Figure 18 , technology can help users to find their perfect fit and achieve one-of-a-kind style by selecting items from the exclusive brands that are hand-selected by an expert stylist. Figure 19 : Case study of Virtual Fitting by DATAGRID. Here, AI is used to generate non-existent photorealistic digital humans, and the company aims to utilise it as a new interface for machines in a future society. This type of work can also be used as an advertising model for apparel e-commerce. More information can be found at COUCHFASHION 38 Virtual Fitting -'Design': Technology can improve customers' online shopping experiences by allowing them to virtually visualise the apparel. Fashion brands and retailers can also use this technology to generate sale or advertising images for their new products without taking new photographs. Virtual Fitting by DATAGRID ( Figure 19 ) uses AI to generate non-existent, photorealistic digital humans. This technology may be utilised as a new interface for machines in a future society where AI has permeated everything. Such work can also be used as an advertising model for apparel e-commerce. Moreover, technology forms an important part of the 'Virtual Retail' business model, which can accelerate the transformation of the fashion industry from selling Products into selling Service. Notably, 2D and 3D versions of this technology have now been developed. Certainly, there are some attempts of utilising neural networks to create virtual fashions which will be transformed into real clothing; e.g., 'Project Muse 45 ' designed by Google in partnership with Zalando in 2016. More information can be found at ARMOI 46 Business Strategy -'Buying': Technology can help businesses to make predictions by using consumer data from various sources and adjusting businesses' selling strategies to drive revenue while maintaining profitability. Retail buyers, planners, and merchandisers can use this type of technology to react faster to market trends, obtain competitive assortment benchmarking, and optimise prices. Technology can analyse data to understand the emotional context behind shoppers' purchases; businesses can then deliver the most relevant, personalised experiences to their consumers. Automatically generating the fashion trending report will be one of the most useful tools. From the above summary, we can see that AI has already been applied for practical use and some conventional fashion business models have been changed. One of the most typical cases is 'fashion searching'. Based on the technology in the group of 'Selling' [4, 18, 115, 59, 105, 68, 131, 41, 56, 161, 88, 165, 103, 92, 143, 43, 74, 35, 12, 111, 40, 71, 188, 39] , online retailers or online shopping platforms can provide their customers with 'intelligent sales person'. Comparing with the basic requirements for the seller, the AI model can recognise the attributes of fashion products and retrieve the related results. This technology undoubtedly affects the business model of online selling. Meanwhile, it is apparent that there is still big room for us to put efforts for improvement, e.g., recommendation of fashion products based on the fashion trends and observations on the customers. Additionally, many technologies have not been applied; e.g., attributes editing [23, 140, 54] in the group of 'Design'. Especially the COVID-19 has changed the conventional practice of the industry. During the global pandemic, many large-scale activities were forced to cancel. Many design houses debuted their new collections via digitalised ways like films or games. For example, as shown in Figure 3 Challenges fAshIon is a challenging field for computer vision. In addition to the most basic function of clothing, i.e., covering the body, well dress-up can bring the feeling of beauty to us. Fashion is an art form, just like music, painting, poetry, etc. Although creating AI models to assist with the different types of works done by fashion practitioners, including sellers, stylists, designers, and buyers can empower the current industry by upgrading its structure while elevating the economy, to teach an AI model to understand fashion is not an easy thing. The first question that should be answered is: 'What is beauty and how to evaluate it?' Philosophers have been discussing this issue for millennia. Different perspectives, such as those of the materialist and the idealist, are juxtaposed and have shaped the meaning of aesthetics. It is generally recognised that beauty produces a feeling that is similar to happiness. The beauty or deformity of an object is caused by its gene. Generally, to perceive beauty, one must perceive the essence that beauty from both internal and external perspectives. Internal meaning differs from external meaning. The external senses may detect traits that do not rely on any priori perception. In other words, if you do not perceive or at least grasp the object, its beauty cannot be perceived. Perception is a major keyword in aesthetics [120] . Haute Couture gowns possess the unique individuality of an object d'art. They are among the last items made by hand, the human hand, whose value is irreplaceable, because it gives its creations that which no machine can ever give: poetry and life. -Christian Dior, 1957 We agree with Dior. No matter how far technology advances, the creative capacity of humans is irreplaceable. In terms of computers, strong AI technology remains a controversial topic, and most mainstream viewpoints oppose conducting related research. Unlike humans, computers cannot express themselves in an emotional context. Creating beautiful and valuable artwork is a complicated challenge for a computer because it has no ability to express emotions. Fortunately, perceiving an object's nature or its structure differs from perceiving its beauty [127] . Therefore, if we only focus on the essential concept, i.e., beauty is balance or harmony among the fundamental elements of fashion in an item of apparel or an outfit, fAshIon tasks can be completed to a certain extent by a computer. Naturally, the basic knowledge learning serves as the foundation of all higher-order fashion tasks. However, even though we do not take 'perception' into consideration, only visually understanding fashion still remains many challenges. Specifically, it is difficult to obtain a clear definition of what constitutes the basic knowledge in fashion; e.g., 'attributes', 'styles', 'trends', and 'design themes'. Here, we will present some easily comprehensible examples as a simple demonstration. The 7 shirt dresses shown in Figure 21 are obviously very different. The attributes that are most immediately obvious are the differences in their colours, branding, and prices. Furthermore, some of the dresses have different prints; the first shirt dress is colour-blocked, whereas the second and the last dresses are striped and the others are solid colours. Moreover, the materials used to make these shirt dresses are also different: the second dress is made of cotton, and the fourth is made of satin. Their silhouettes also differ: the fourth dress is H-line and the sixth is A-line. Thus far, we have only considered the main differences. If we investigate in further detail, we find that the waists of these 7 dresses are all different from each other. The design of the pockets and the opening designs are also different. There are many aspects that make these shirt dresses different from each other. We have only described a few such aspects here. If we imagine these design attributes as letters in an alphabet, then the garment is a word made up of these letters. Then, an outfit is like a sentence composed of many words, all obeying a fixed grammar. Many outfits gather together to express a certain theme, such as a collection in a fashion show, tell a story composed of many sentences. Now, we ask two questions: • Is the love of Romeo and Juliet related to the 26 letters of the Roman alphabet? • Is Hamlet 65 defined as a tragedy only because of some of the words or sentences used in the play? Here, we use the show S. W. A. L. K. II 66 to explore similar concepts. The link below presents the related media and information that can be easily find, such as a paragraph describing this show: The connectivity of interdependence becomes revalued in times of separation. The reliance of one person upon another is a vital pas de deux activated by instinct and trust. For the Co-Ed Collection Spring-Summer 2021, Maison Margiela interprets this concord through the tango. Vigorous and intense, the dance is cathartic: releasing the spirit of the old, it inspires the lust to move on. It compels acceptance; it heralds new beginnings; it beckons change. We present some of the looks in this series in Figure 22 . How did the designer use tulle, velvet, worn, silk, chiffon, leather, and other materials to express the theme of this show, as mentioned above? • Are the feelings or emotions evoked by this collection closely related to the attributes of fashion that have been adopted? • Can the theme expressed by this fashion show be understood by considering only Look 29 (a white stretch-tulle circular cut dress with farmed plumed trims worn over a white muslin, circular cut long-sleeved dress with white chiffon ghillies and white painted leather Tango pumps) or Look 30 (an ecru wool, double-breasted tuxedo with silk duchesse lapels worn over a white tuxedo shirt with white painted leather Hyperion ankle-strap shoes)? As described in the summary of the hierarchy of visual understanding 67 : the first, or lowest, level in the hierarchy is data (discrete elements), the second level is information (linked elements), the third level is knowledge (organised information), and the highest level is wisdom (applied knowledge). The challenges increase progressively at higher levels of visual understanding, and may include: • Recognising data (visual recognition of fashion images). • Turning data into effective information (obtaining fashion-related information from the recognised attributes, e.g., recognising a style or evaluating an outfit and knowing the behind reason for this evaluation judgement). • Translating information into knowledge based on understanding (understanding clothing aesthetics, knowing how to predict trends and why trends occur). • Applying information freely and forming insights (knowing how to create beauty and lead trends). fAshIon is a fresh field that full of opportunities. As summarised in the last Section, there has long way to go, even if only visually understand fashion. Undoubtedly, large amount of problems remain for researchers to tackle. In additon, under the impact of the epidemic, embracing technology becoming the new trend in fashion. The whole industry will be unavoidably be reshaped in the Post-COVID-19 Era. Addition to the Balenciage Fall 2021, many other designers also expressed their own thinkings to the fAshIon in the near future; e.g., after Prada 2021 SS 68 , there has 'a conversation' in which Miuccia talked her ideas about the relationship between humans and machines. We do not forecast the future of fAshIon in the industry due to our limited abilities. Here we would like to use what Demna Gvasalia said to represent our viewpoints: 'I believe in a future that is spiritual. Loading a forgotten past.' In the following, we conduct a focused analysis of opportunities for research and summarise the possible directions: Opportunities in 'Evaluation': 'In God we trust; all others must bring data 69 '. As mentioned before, very few research have focused on the evaluation protocols in fAshIon. In addition to the requirements for the algorithms proposed for a specific task to be compared fairly, a company also needs clear criteria to evaluate whether a model is qualified to be embedded in their online products or whether the modelled numbers or predictions can be trusted. Opportunities in 'Basic Tech': Even though the research in the group of 'Basic Tech' are not directly related to a role in fashion, it is the foundation for computer to understand fashion images. As mentioned before, most studies targeted on clothing parsing [86, 83, 84, 26, 99, 163, 25, 94, 162, 158, 85, 100, 77, 87, 23, 42, 97, 24] . However, in terms of clothing parsing, none of them paid attention to the segmentation between the front piece and the back piece of a given single clothing item in a fashion image. Meanwhile, when several garments are dressed by a person at the same time, it arises a occlusion problem. In addition, it is great help if the model can segment the garment according to its design regions; e.g., neck region, sleeve region, body region, cuff region, shoulder region, chest region, waist region etc. Opportunities in 'Selling': From Figure 22 , it is easy to see that there are many attributes of fashion that should be recognised. This is a multi-label classification task. We roughly estimate that the number of classification labels is over 250 [188] . Moreover, there is no existing open-access datasets of comprehensive and well-annotated fashion attributes. Thus, the efforts to train an AI model using current data will face challenges such as weakly labelled data and noises. The diverse format of fashion image data, including fashion show images, product images, model images, street photos, and social media images, also increases the difficulty of recognition. Here we would like to recall the basic requirements for the seller: 1. can recognise the detailed attributes of fashion products and know how to find similar items according to a general description using simple words provided by his or her customers; 2. can accurately recommend fashion products according to observations on both his or her customers and of trends; 3. can reply quickly to his or her customers about whether a particular item or size is available or out of stock. Therefore, the conclude the possible future directions including: 1. multi-label classification in weakly supervised manner; 2. fine-grained attributes recognition in webly data or videos; 3. image retrieval in massive data. Opportunities in 'Styling': Current research in the group of 'Styling' are mainly focused on creating outfit composition for online recommendation. The basic requirements for a stylist are: 1. has a good sense of clothing aesthetics, such as a sense of colour, a sense of texture, and a sense of silhouette, and can easily create a well-composed outfit; 2. knows how to make an outfit attain visual balance for a given customer and mix and match according to different situations, such as for an occasion, the seasons, or body figure; 3. understands what beauty is and can convincingly explain his or her selections, enabling him or her to persuade the customers and provide satisfactory service; 4. always stays aware of fashion trends and can apply these trends to his or her work. Therefore, we conclude the possible future directions including: 1. evaluating an outfit is good or not with convicing explanation; 2. understanding about the basic 'principles' of mix and match and behind logic; 3. recommending personalised styling advice and catching up the trend in time. Opportunities in 'Design': Most of current works focus on helping designers or online retailers to visually demonstrate their ideas, e.g., attribute editing [23, 140, 54] , 3D garment generation [118, 5, 169, 150, 37, 38] , virtual try-on [125, 21, 72, 20, 13, 45] , etc. Certainly, with a holistic view of the 'Design' research, there remain many interesting but challenging tasks. Can AI models create new designs based on a theme with different predetermined constraints, such as seasons, trends, brand image, etc.? Thus, we conclude the possible future directions including: 1. utilising generation methods in virtual marketing; 2. appling generation methods in the process of fashion design; 3. expressing a concept or idea using fashion language. Opportunities in 'Buying': Most research in this area focus on fashion forecasting, prediction, and selling [53, 110, 1, 144, 124, 109, 151, 142, 132] . As a good buyer, he or she should: 1. can draw conclusions from records of previous sales; 2. have great insight into the market and the ability to make predictions; 3. very experienced in fashion retail and fully knows the market trends. Therefore, we conclude the possible future directions including: 1. predicting the market trends based on the analysis of massive data; 2. understanding the culture and inferring the effect of an event on fashion which can lead the trend of fashion. 'Big data can tell us what is wrong, not what is right 70 '. Therefore, we should be very sceptical of any 'big data analyst' or 'data scientist' who claims to be able to explain a system in a particular domain without the requisite domain expertise or intimate knowledge of the underlying system under consideration 71 . To really help the industry, it is better to design tasks and solutions based on the domain knowledge. Modeling fashion influence from photos From paris to berlin: Discovering fashion style influences around the world Fashion outfit generation for e-commerce Apparel classification with style Multi-view consistency loss for improved single-image 3d reconstruction of clothed people Segmentation task for fashion and apparel Mode regularized generative adversarial networks Describing clothing by semantic attributes Tailorgan: Making user-defined fashion designs Pog: personalized outfit generation for fashion recommendation at alibaba ifashion Fashion meets computer vision: A survey Video2shop: Exact matching clothes in videos to online shopping images Pivtons: Pose invariant virtual try-on shoe with conditional image completion Context-aware visual compatibility prediction Fashiongan: Display your fashion design using conditional generative adversarial nets Dressing as a whole: Outfit compatibility learning based on node-wise graph neural networks Fashioning with networks: neural style transfer to design clothes Style finder: Fine-grained clothing style detection and retrieval Soft-gated warping-gan for pose-guided person image synthesis Towards multi-pose guided virtual try-on network Fw-gan: Flow-navigated warping gan for video virtual try-on Fashion editing with multi-scale attention normalization Fashion editing with adversarial parsing learning A deformable mixture parsing model with parselets Towards unified human parsing and pose estimation Parsing based on parselets: A unified deformable mixture model for human parsing Class rectification hard mining for imbalanced deep learning Multi-task curriculum transfer deep learning of clothing attributes Imbalanced deep learning by minority class incremental rectification A variational u-net for conditional appearance and shape generation Deepfashion2: A versatile benchmark for detection, pose estimation, segmentation and re-identification of clothing images Aesthetics, personalization and recommendation: A survey on deep learning in fashion Ladn: Local adversarial disentangling network for facial makeup and de-makeup Understanding fashion trends from street photos via neighbor-constrained embedding learning Multimodal and multi-domain embedding learning for fashion retrieval and analysis Garnet: A two-stream network for fast and accurate 3d cloth draping Garnet++: Improving fast and accurate static 3d cloth draping by curvature loss The imaterialist fashion attribute dataset Brand> logo: Visual analysis of fashion brands Where to buy it: Matching street clothing photos in online shops Bottom-up/top-down image parsing with attribute grammar Automatic spatially-aware fashion concept discovery Learning fashion compatibility with bidirectional lstms Viton: An image-based virtual try-on network Clothflow: A flow-based model for clothed person generation Finet: Compatible and diverse fashion image inpainting From design draft to real attire: Unaligned fashion image translation Attr2style: A transfer learning approach for inferring fashion styles via apparel attributes. arXiv e-prints Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering Learning compatibility across categories for heterogeneous item recommendation Dress with style: Learning style from joint deep embedding of clothing styles and body shapes From culture to clothing: Discovering the world events behind a century of fashion images Fashion++: Minimal edits for outfit improvement Deep generative models with learnable knowledge constraints Cross-domain image retrieval with a dual attribute-aware ranking network Arch: Animatable reconstruction of clothed humans Multi-Label Fashion Image Classification with Minimal Human Supervision Large scale visual recommendations from street fashion images The fashionpedia ontology and fashion segmentation dataset Bcnet: Learning body and cloth shape from a single image Fashion style generator Visually-aware fashion recommendation and design with generative image models Complete the look: Scene-based complementary product recommendation An overview of image recognition and retrieval of clothing items Deepwear: a case study of collaborative design between human and artificial intelligence Gans-based clothes design: Pattern maker is all you need to design clothing Hipster wars: Discovering elements of fashion styles Self-supervised visual attribute learning for fashion compatibility Whittlesearch: Image search with relative attribute feedback Ontology-driven hierarchical deep learning for fashion recognition Shineon: Illuminating design choices for practical video-based virtual clothing try-on Attention-based fusion for outfit recommendation Cross-modal search for fashion attributes A global-local embedding module for fashion landmark detection Coherent and controllable outfit generation Holistic, instance-level human parsing Beautygan: Instance-level facial makeup transfer with deep generative adversarial network Hierarchical fashion graph network for personalized outfit recommendation A deep learning based interactive sketching system for fashion images design Spatial-aware non-local attention for fashion landmark detection Deep physics-aware inference of cloth deformation for monocular human performance capture Deep human parsing with active template regression Human parsing with contextualized convolutional neural network Clothes co-parsing via joint image segmentation and labeling with application to clothing retrieval Semantic object parsing with graph lstm Look into person: Joint body parsing & pose estimation network and a new benchmark Rapid clothing retrieval via deep learning of binary codes and hierarchical search Aggregation and finetuning for clothes landmark detection Fashion outfit complementary item retrieval Learning personal tastes in choosing fashion outfits Mvc: A dataset for view-invariant clothing retrieval and attribute prediction Toward ai fashion design: An attribute-gan model for clothing match Weakly supervised graph propagation towards collective image parsing Hi, magic closet, tell me what to wear! Street-to-shop: Cross-scenario clothing retrieval via parts alignment and auxiliary set Fashion parsing with weak color-category labels Fashion analysis: Current techniques and future directions Fashion parsing with video context Surveillance video parsing with single frame supervision Learning diverse fashion collocation by neural graph filtering Swapgan: A multistage generative approach for person-toperson fashion style transfer Deepfashion: Powering robust clothes recognition and retrieval with rich annotations Fashion landmark detection in the wild Fashion 10000: an enriched social image dataset for fashion and clothing Toward better reconstruction of style images with gans. KDDW on ML meets fashion Disentangled person image generation Yixin Cao, and Tat-Seng Chua. Who, where, and what to wear? extracting fashion knowledge from social media Knowledge enhanced neural fashion trend forecasting Discovering underground maps from fashion Tiered deep similarity search for fashion p05953139/AI-in-Fashion-Market-Research-Report-by-Product-Type-by-Component-by-Deployment-by-Applicat html? StreetStyle: Exploring world-wide clothing styles from millions of photos Image-based recommendations on styles and substitutes Snapper: fashion coordinate image retrieval system Siclope: Silhouette-based clothed people Poly-gan: Multi-conditioned gan for fashion synthesis Tailornet: Predicting clothing in 3d as a function of human pose, shape and garment style Learning fashion compatibility across apparel categories for outfit recommendation Essays on the intellectual powers of man. Number 66 Interactive systems for fashion clothing recommendation Pai-bpr: Personalized outfit recommendation scheme with attribute-wise interpretability Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization Explainable ai based interventions for pre-season decision making in fashion retail Learning-based animation of clothing for virtual try-on Design: Design inspiration from generative networks The concept of the aesthetic Assessing fashion recommendations: A multifaceted offline evaluation approach Learning to synthesize fashion textures Deformable gans for posebased human image generation Neuroaesthetics in Fashion: Modeling the Perception of Fashionability Nilpa Jha, and Aruna Rajan. Fashion retail: Forecasting demand for new items Towards a unified framework for visual compatibility prediction When multimedia meets fashion Neurostylist: Neural compatibility modeling for clothing matching Neural compatibility modeling with attentive knowledge distillation Gp-bpr: Personalized compatibility modeling for clothing matching Personalised fashion recommendation using deep learning Cfrs: A trends-driven collaborative fashion recommendation system Deepcloth: Neural garment representation for shape and style editing Learning fashion compatibility across categories with deep multimodal neural networks Breaking moravec's paradox: Visual-based distribution in smart fashion retail What makes a style: Experimental analysis of fashion prediction Cat street: Chronicle archive of tokyo street-fashion Anime-to-real clothing: Cosplay costume generation via image-to-image translation Toward explainable fashion recommendation Looking at outfit to parse clothing Recommending outfits from personal closet Improving apparel detection with category grouping and multi-grained branches Bodynet: Volumetric inference of 3d human body shapes Product age based demand forecast model for fashion retail Learning type-aware embeddings for fashion compatibility Addressing the cold-start problem in outfit recommendation using visual preference modelling Automatic attribute discovery with neural activations Who blocks who: Simultaneous clothing segmentation for grouping images Outfit compatibility prediction and diagnosis with multi-layered comparison network Example-based real-time clothing synthesis for virtual agents Enhanced reweighted mrfs for efficient fashion image parsing Joint multi-person pose estimation and semantic part segmentation Texturegan: Controlling deep image synthesis with texture patches Learning from massive noisy labeled data for image classification Parsing clothing in fashion photographs Paper doll parsing: Retrieving similar styles to parse clothing items Retrieving similar styles to parse clothing Mix and match: Joint model for clothing and attribute recognition Unconstrained fashion landmark detection via hierarchical recurrent transformer networks Interpretable fashion matching with rich attributes Learning to match on graph for fashion compatibility modeling Neural 3d clothes retargeting from a single image Fine-grained visual comparisons with local learning Fine-grained visual comparisons with local learning Personalized fashion design Layout-graph reasoning for fashion landmark detection Garment design with generative adversarial networks Cross-domain shoe retrieval with a semantic hierarchy of attribute classification network Street-to-shop shoe retrieval with multi-scale viewpoint invariant triplet network Deepshoe: An improved multi-task viewinvariant cnn for street-to-shop shoe retrieval Pose-normalized and appearance-preserved street-to-shop clothing image generation and feature learning Clothes collocation recommendations by compatibility learning Learning color compatibility in fashion outfits Deep detail enhancement for any garment A compensation method of two-stage image generation for humanai collaborated in-situ fashion design in augmented reality environment Personalized fashion recommendation from personal social media data: An item-to-set metric learning approach Text guided person image synthesis Be your own prada: Fashion synthesis with structural coherence Fashion landmark detection and category classification for robotics Fashion meets ai technology Fashionai: A hierarchical dataset for fashion understanding Regularizing reasons for outfit evaluation with gradient penalty This research was funded by the Laboratory for Artificial Intelligence in Design (Project Code: RP3-1), Hong Kong Special Administrative Region. The title, fAshIon after fashion, was inspired by a seminar given by Professor Hazel Clark at Parsons School of Design to accompany the excellent exhibition fashion after Fashion. The authors would also like to express deep appreciation to Professor Zowie Broach at the Royal College of Art for her enlightening comments and inspirational contributions from the very beginning of the journey of fAshIon. 70 N. Taleb 71 https://medium.com/@adambreckler/in-god-we-trust-all-others-bring-data-96784d01e9be