THE ROBOFASHION MAKEOVER

While hu­man in­spi­ra­tion will al­ways be the pri­mary source of new de­signs, ma­chines are learn­ing to spot cur­rent trends and pre­dict new ones. Sylvia Chang re­ports from Hong Kong.

China Daily (Latin America Weekly) - - China -

When Thai­land’s King Bhu­mi­bol Adulyadej died in Oc­to­ber 2016, a low- to mid-end Hong Kong fash­ion brand sold out of black shirts and white ones — the col­ors of re­spect worn dur­ing the of­fi­cial mourn­ing pe­riod — al­most overnight.

That prompted de­sign­ers and mer­chan­dis­ers to be­gin won­der­ing whether such color-fo­cused shop­ping sprees could be pre­dicted.

Tra­di­tion­ally, in­dus­try ex­perts pre­dict fash­ion trends based on their own ex­pe­ri­ences and sense of style. These days, how­ever, so­cial me­dia’s im­pact on con­sumers’ be­hav­ior comes into play, and a fash­ion in­dus­try rev­o­lu­tion is un­der­way, with help from ar­ti­fi­cial in­tel­li­gence tech­nolo­gies.

“We in­tend to pro­vide a real-time es­ti­ma­tion of fash­ion color trends by us­ing big data from so­cial me­dia,” said Glo­ria Yao, direc­tor of project de­vel­op­ment at the Hong Kong Re­search In­sti­tute of Tex­tiles and Ap­parel, who is work­ing with a team that is de­vel­op­ing an AI-based col­or­pre­dic­tion model.

Yao said sales pre­dic­tion is very im­por­tant for fash­ion busi­nesses as it makes for ef­fec­tive in­ven­tory man­age­ment, which in­creases prof­its by re­duc­ing mark­downs due to over­stock­ing.

So­cial me­dia im­pact

Ac­cord­ing to a 2016 re­port by ad­ver­tis­ing agency PMX, which is based in the United States, so­cial me­dia posts drive 6.3 per­cent of web­site traf­fic to lux­ury brands, with Face­book the plat­form that drives the most sales.

That prompted Yao and her team to fo­cus on hun­dreds of in­flu­en­tial fash­ion in­dus­try ac­counts on Face­book, in­clud­ing over 500 fash­ion brands, 100 fash­ion mag­a­zines, 150 de­sign­ers and 500 so­cial me­dia celebri­ties. To qual­ify as in­flu­en­tial, Yao said, a brand should have at least 100,000 fol­low­ers, a mag­a­zine 10,000 fol­low­ers, and a de­signer or celebrity 1,000 fol­low­ers. In ad­di­tion, the ac­counts should be up­dated fre­quently and be able to spark com­mu­nity con­ver­sa­tions.

“The aim is to cap­ture the on­go­ing so­cial and cul­tural events from so­cial me­dia posts and val­i­date their im­pact on con­sumers’ color pref­er­ences,” Yao said.

An AI tech­nol­ogy known as nat­u­ral lan­guage pro­cess­ing in­ter­prets the text and im­ages in Face­book posts and helps to iden­tify “authen­tic fash­ion posts” re­lated to color.

“If a celebrity says, ‘I’m walk­ing on a red car­pet at a film fes­ti­val,’ the ‘red’ here has noth­ing to do with fash­ion,” Yao ex­plained. “But if she says, ‘I like my red-col­ored dress to­day,’ then this ‘red’ is the tar­get of our data col­lec­tion.”

The in­sti­tute works with a well-known Hong Kong fash­ion brand that pro­vides its data on sales, in­ven­to­ries, prices, shop lo­ca­tions and mar­ket­ing in­for­ma­tion.

Yao’s team found a cor­re­la­tion when com­par­ing data from Face­book posts with that from the brand, with a color trend start­ing on Face­book ahead of an uptick in sales.

“If we can de­ter­mine the time gap be­tween the two trends and their re­la­tion­ship, we will have more ac­cu­rate color pre­dic­tions,” Yao said. His­tor­i­cal data re­veal an av­er­age ac­cu­racy of about 70 per­cent, but this may im­prove as data are up­dated more of­ten.

The project will fin­ish in April, when a pro­duc­tion li­cense will be in­tro­duced and com­mer­cial­ized for ap­pli­ca­tion in the in­dus­try.

At present, Yao said, the color-pre­dic­tion model fits dif­fer­ent brands, “but with some ad­just­ments”. The rea­son is that dif­fer­ent brands may tar­get con­sumers with vary­ing color pref­er­ences.

Ap­ply­ing AI to makeup

More fash­ion brands are ap­ply­ing AI tech­nolo­gies. Cos­met­ics gi­ant L’Oreal has a smart­phone app that al­lows cus­tomers to “try on” lip­sticks, eye shadow and eye­lash styles. Lux­ury brands like Burberry give shop­pers in­tel­li­gent ad­vice on which fash­ion items suit them best.

“AI en­hance­ments will go be­yond the tra­di­tional ar­eas of ma­chine tasks into cre­ative and cus­tomer-in­ter­ac­tion pro­cesses, blur­ring the line be­tween tech­nol­ogy and cre­ativ­ity,” US man­age­ment con­sult­ing firm McKin­sey & Com­pany said in a re­port on fash­ion trends this year.

Fash­ion is about de­sign and the cus­tomer’s emo­tions and feel­ings — com­bin­ing phys­i­cal ma­te­ri­als with hu­man senses. So how can it be read by a ma­chine?

The first step, said Calvin Wong, a pro­fes­sor at Hong Kong Polytech­nic Univer­sity’s In­sti­tute of Tex­tiles and Cloth­ing, is to teach the ma­chine to an­a­lyze a fash­ion image.

That en­sures “the ma­chine un­der­stands the fash­ion world”, said Wong, who is lead­ing a team of a dozen re­searchers that is work­ing with Alibaba Group to en­hance the search func­tions on its Taobao e-com­merce plat­form.

The big­gest chal­lenge fac­ing Taobao is the sheer chaos of prod­ucts that have not been prop­erly tagged, said Jia Men­glei, a se­nior en­gi­neer at Alibaba, with cus­tomers com­plain­ing that it’s hard to find what they’re look­ing for.

Li Yuan, 26, and a keen on­line shop­per, said: “There are just tons of items on Taobao. I of­ten have to spend hours on it se­lect­ing what I need.

“Some­times it makes me feel dizzy.”

Jia said us­ing a data set of fash­ion at­tributes pro­duced by Wong’s team will make tag­ging more ex­act and pre­dictable, al­low­ing Alibaba to “re­build the world map of con­sumer goods”.

Un­der­stand­ing im­ages

Ap­parel comes in in­nu­mer­able pat­terns, each with dif­fer­ent char­ac­ter­is­tics and el­e­ments that de­fine it as “fash­ion”. The va­ri­ety of ma­te­ri­als, col­ors, shapes and styles cre­ates skirts, pants, T-shirts, dresses, sweaters and other items that ap­peal to a range of tastes.

Take a sweater. It can have a pu­ri­tan col­lar, a shirt col­lar, a rib col­lar — the list goes on. To train a ma­chine to un­der­stand a sweater, you need to in­stall a data set of im­ages of all kinds of sweaters, tagged with their de­sign at­tributes. The AI net­work, or a sin­gle ma­chine trained to have pre­dic­tive abil­ity, would then rec­og­nize a spe­cific type of sweater the next time it sees one.

The data set is most im­por­tant. It serves as a guide­line for the ma­chine’s learn­ing process and sets the foun­da­tion for all fol­low-up re­search.

“It needs to be pro­fes­sional enough to meet the re­quire­ments of ma­chine train­ing,” Wong said, in­clud­ing com­plete and pre­cise at­tributes for each style that may be ap­plied to the ba­sic image.

“For a sweater with a turtle­neck, you need to give it a tag both as a sweater and a turtle­neck. If it’s iden­ti­fied some­times as a sweater and some­times as a turtle­neck, the ma­chine will be con­fused.”

Wong said the ac­cu­racy of searches for cloth­ing im­ages on the in­ter­net is only about 50 per­cent, some­thing he at­tributes to the lack of suf­fi­cient pro­fes­sional data.

The AI sys­tem adopted by Alibaba has been trained to rec­og­nize more than 500,000 out­fits.

The se­cond level of AI ap­pli­ca­tion in fash­ion is to an­a­lyze cur­rent styles, eval­u­ate their most pop­u­lar el­e­ments, pro­vide rec­om­men­da­tions and even pre­dict trends.

Wong said that if a ma­chine has a data set of the lat­est im­ages of out­fits in the fash­ion world, it could cal­cu­late which el­e­ments are emerg­ing as pop­u­lar and may be­come the trend of the fore­see­able fu­ture. Then it could of­fer rec­om­men­da­tions for con­sumers, based on pre­dic­tive anal­y­sis and in­di­vid­ual pref­er­ences.

The team ex­pects that in the near fu­ture AI will be able to col­lect more data from cus­tomers, pro­vid­ing im­por­tant in­for­ma­tion about the pop­u­lar­ity of cer­tain styles. The data can note the amount of time a cus­tomer spends look­ing at a par­tic­u­lar piece of cloth­ing us­ing a “smart mir­ror” and whether they de­cide to buy it, along with other as­pects of the fash­ion ex­pe­ri­ence.

Alibaba will com­bine data from Taobao with that col­lected in off­line shops, aim­ing to pro­vide an all-around data ser­vice for the in­dus­try — from fash­ion de­sign and ma­te­rial se­lec­tion to stor­age and lo­gis­tics.

AI ap­pears set to trans­form the fash­ion in­dus­try, but is there any­thing a ma­chine is not likely to mas­ter in the near fu­ture?

Yes, said Wong, be­cause, in the end, de­sign­ers are ir­re­place­able. The core of de­sign is the in­spi­ra­tion, he said, and a ma­chine can only serve as an as­sis­tant, at best.

Con­tact the writer at [email protected]­nadai­lyhk.com

PRO­VIDED TO CHINA DAILY

A model uses AI tech­nol­ogy to try on clothes at a fash­ion con­cept store at Hong Kong Polytech­nic Univer­sity.

ZHU XINGXIN / CHINA DAILY

Vis­i­tors ex­pe­ri­encesmart fit­ting at Alibaba’s booth at an ex­hi­bi­tion in Fuzhou, Fu­jian prov­ince, in April.

Newspapers in English

Newspapers from Argentina

© PressReader. All rights reserved.