7 ms·

Building a fashion search engine with deep learning

82 points by houqp 8y ago

ceejayoz 8y ago
I'd like to see someone tackle sizing. A 4X dress on Amazon will often wind up being the equivalent of an American medium because it's from a Chinese vendor. It's baffling to me that Amazon doesn't make their clothing vendors enter a basic bust/waist/hip set of measurements and let people search by those.
- egman_ekki 8y ago
  I think e.g. easysize.me are trying to solve that.
  - gkhusainova 8y ago
    Thanks for the shoutout! Gulnaz here from Easysize. We use a range of factors (combined with ML) to recommend the right size and fit: from gathering product-specific post-purchase feedback, analyzing orders & returns and asking shoppers about their unique sizing challenges. IMHO when it comes to recommending the right size at scale, the key challenges are: 1. taking into account fit and style: the same shoppers may like slim-fitted shirts and oversized sweaters. 2. keeping UX super simple: shoppers get bored and annoyed when they're asked too many questions or asked to perform some other actions. 3. taking into account unique product features: fabric, clothing cut etc.
- dman 8y ago
  Wish there was some kind of international standard for sizing - where toddlers / kids / adults lived within the same namespace. As a recent parent I have been going crazy seeing the inconsitent sizing labels for toddlers / children.
- hcrisp 8y ago
  Outdoor clothier PrAna has already done this. They ask for your height, weight, age, and what size of what brand fits you the best to predict best match. ML at work. https://www.prana.com/true-fit https://www.prana.com/true-fit
  - tomnipotent 8y ago
    None of it works. I'm in the fashion e-com space and returns/fit is a serious issue, but no one has really made any progress when ordering online. Even services like TrueFit/FitFinder barely make a dent.
    - ricberw 8y ago
      I’m working with my team on a solution that might be slightly different from what you’ve seen thus far — check out Drapr.com or download our app on the App Store (iOS only for now). Photorealistic, physically accurate virtual try-ons.
    - mendeza 8y ago
      Do you think if cameras can capture real depth values, or capture volumetric information, could you at least prevent shoppers buying clothes that will not likely fit. I don’t think personalizes fit would be solved, but if you could measure dimensions and include that in recommendation engines, it could provide more insight.
      - tomnipotent 8y ago
        Still too many features to make a difference - all we end up doing is shifting one distribution for another, and the net result is the same number of unsatisfied customers but just for different reasons. Neck & shoulders, length of waist vs. legs, arm length or face shape. All of the tech I've seen in this space does great work for improving conversion rate - which I take advantage of myself - but overall there has been no noticeable impact on net sales.
- web007 8y ago
  I worked on this problem a few gigs back at Polyvore! tl;dr: It's nearly impossible to solve. As you may already know, "Standard Sizing" is a misnomer. There is no single standard for a size. US vs EU vs JP vs other are all the same number (size 4) but different nominal sizes. Add to that the chaos of standard sizes between brands (brand A size 4 != brand B size 4 != brand C size 4) and even within the same brand (brand A product 1 size 4 != brand A product 2 size 4) and you find that there's no single answer to matching a human to a size. Even your suggestion of measurements won't help without more data around it. Clothing shape and fabric type matter nearly as much as the measurements, since one may be a form-fitting 90cm and another may be a breezy and stretchy 90cm. Personal preference matters as well, so you may prefer breezy while I prefer form-fitting.
  - ceejayoz 8y ago
    So I get that it's likely to be impossible to fully solve anytime soon, but I'd think at least requiring a couple of basic measurements would make a dent. I've seen exactly what you describe - sizes varying between countries, between brands, between individual items in a brand. A lot of the listings on Amazon include a custom size chart now, but it's not semantic data - it's just a screenshot in the product listing. Example: https://www.amazon.com/Beautifulfashionlife-Waisted-Pleated-Cosplay-Costumes/dp/B01179J4S6/ https://www.amazon.com/Beautifulfashionlife-Waisted-Pleated-... This one lists a XXXX-Large as a 34" waist, and a Large as a 28" waist. It shows up if you search Amazon for "plus size skirt"; if Amazon made vendors put in the waist measurements and let them be filterable based on that, someone in a US size 18 dress wouldn't be offered this as an option. Again, I get the difference between a 90cm loose fit and a 90cm tight fit. I should be able to tell a major ecommerce site "look, I'm 150cm... fuck off with the 90cm stuff".
    - twairball 8y ago
      MeasureAI performs clothing measurements from a photo, no tape measure required. Link: https://measure.productai.com https://measure.productai.com Disclaimer: I worked on this product. Standardization is one way to go, but for us we quickly found that with clothing styles, edge cases are the norm, not the exception.
- doctorpangloss 8y ago
  An alternative interpretation of the data is that people are buying clothes on impulse, so it doesn't really matter if the clothes fit or not. Surely the graveyard of doing-it-smarter retailers lend some credence to the theory. You're not the first person nor the last to want clothes to fit you well. But you just don't buy new clothes all the time like the impulsive crazies do, and they're the "whales" of the fashion industry.
whatrocks 8y ago
This similar clothing recommender seems like something that Netflix or Amazon Prime will add to their pause screens on their original content - letting viewers purchase similar things to what the characters are wearing right on screen. I remember searching for years for some dumb shirt that Owen Wilson wore in some dumb movie, but I was so happy when I finally found something like it.
- thesehands 8y ago
  That is how ASOS started, As Seen On Screen. Now doing over 2 billion in turnover a year
  - whatrocks 8y ago
    Not surprising at all! Instagram is probably next for this, then. Every single non-ad post could be a way to sell you the clothes or widgets that your friends are wearing/using in their fabulous lives.
- esturk 8y ago
  Better be opt-in or it would be annoying.
petard 8y ago
Working in this area the biggest challenge is getting the details right. Small difference like cut-outs on a dress make a huge difference in terms of end-user relevance. Same for colors, millennial orange comes in many shades. Deep Learning is pretty bad with such details.

82 points by houqp 8y ago

I'd like to see someone tackle sizing. A 4X dress on Amazon will often wind up being the equivalent of an American medium because it's from a Chinese vendor. It's baffling to me that Amazon doesn't make their clothing vendors enter a basic bust/waist/hip set of measurements and let people search by those.

egman_ekki 8y ago
I think e.g. easysize.me are trying to solve that.
- gkhusainova 8y ago
  Thanks for the shoutout! Gulnaz here from Easysize. We use a range of factors (combined with ML) to recommend the right size and fit: from gathering product-specific post-purchase feedback, analyzing orders & returns and asking shoppers about their unique sizing challenges. IMHO when it comes to recommending the right size at scale, the key challenges are: 1. taking into account fit and style: the same shoppers may like slim-fitted shirts and oversized sweaters. 2. keeping UX super simple: shoppers get bored and annoyed when they're asked too many questions or asked to perform some other actions. 3. taking into account unique product features: fabric, clothing cut etc.
dman 8y ago
Wish there was some kind of international standard for sizing - where toddlers / kids / adults lived within the same namespace. As a recent parent I have been going crazy seeing the inconsitent sizing labels for toddlers / children.
hcrisp 8y ago
Outdoor clothier PrAna has already done this. They ask for your height, weight, age, and what size of what brand fits you the best to predict best match. ML at work. https://www.prana.com/true-fit https://www.prana.com/true-fit
- tomnipotent 8y ago
  None of it works. I'm in the fashion e-com space and returns/fit is a serious issue, but no one has really made any progress when ordering online. Even services like TrueFit/FitFinder barely make a dent.
  - ricberw 8y ago
    I’m working with my team on a solution that might be slightly different from what you’ve seen thus far — check out Drapr.com or download our app on the App Store (iOS only for now). Photorealistic, physically accurate virtual try-ons.
  - mendeza 8y ago
    Do you think if cameras can capture real depth values, or capture volumetric information, could you at least prevent shoppers buying clothes that will not likely fit. I don’t think personalizes fit would be solved, but if you could measure dimensions and include that in recommendation engines, it could provide more insight.
    - tomnipotent 8y ago
      Still too many features to make a difference - all we end up doing is shifting one distribution for another, and the net result is the same number of unsatisfied customers but just for different reasons. Neck & shoulders, length of waist vs. legs, arm length or face shape. All of the tech I've seen in this space does great work for improving conversion rate - which I take advantage of myself - but overall there has been no noticeable impact on net sales.
web007 8y ago
I worked on this problem a few gigs back at Polyvore! tl;dr: It's nearly impossible to solve. As you may already know, "Standard Sizing" is a misnomer. There is no single standard for a size. US vs EU vs JP vs other are all the same number (size 4) but different nominal sizes. Add to that the chaos of standard sizes between brands (brand A size 4 != brand B size 4 != brand C size 4) and even within the same brand (brand A product 1 size 4 != brand A product 2 size 4) and you find that there's no single answer to matching a human to a size. Even your suggestion of measurements won't help without more data around it. Clothing shape and fabric type matter nearly as much as the measurements, since one may be a form-fitting 90cm and another may be a breezy and stretchy 90cm. Personal preference matters as well, so you may prefer breezy while I prefer form-fitting.
- ceejayoz 8y ago
  So I get that it's likely to be impossible to fully solve anytime soon, but I'd think at least requiring a couple of basic measurements would make a dent. I've seen exactly what you describe - sizes varying between countries, between brands, between individual items in a brand. A lot of the listings on Amazon include a custom size chart now, but it's not semantic data - it's just a screenshot in the product listing. Example: https://www.amazon.com/Beautifulfashionlife-Waisted-Pleated-Cosplay-Costumes/dp/B01179J4S6/ https://www.amazon.com/Beautifulfashionlife-Waisted-Pleated-... This one lists a XXXX-Large as a 34" waist, and a Large as a 28" waist. It shows up if you search Amazon for "plus size skirt"; if Amazon made vendors put in the waist measurements and let them be filterable based on that, someone in a US size 18 dress wouldn't be offered this as an option. Again, I get the difference between a 90cm loose fit and a 90cm tight fit. I should be able to tell a major ecommerce site "look, I'm 150cm... fuck off with the 90cm stuff".
  - twairball 8y ago
    MeasureAI performs clothing measurements from a photo, no tape measure required. Link: https://measure.productai.com https://measure.productai.com Disclaimer: I worked on this product. Standardization is one way to go, but for us we quickly found that with clothing styles, edge cases are the norm, not the exception.
doctorpangloss 8y ago
An alternative interpretation of the data is that people are buying clothes on impulse, so it doesn't really matter if the clothes fit or not. Surely the graveyard of doing-it-smarter retailers lend some credence to the theory. You're not the first person nor the last to want clothes to fit you well. But you just don't buy new clothes all the time like the impulsive crazies do, and they're the "whales" of the fashion industry.

whatrocks 8y ago

This similar clothing recommender seems like something that Netflix or Amazon Prime will add to their pause screens on their original content - letting viewers purchase similar things to what the characters are wearing right on screen. I remember searching for years for some dumb shirt that Owen Wilson wore in some dumb movie, but I was so happy when I finally found something like it.

thesehands 8y ago
That is how ASOS started, As Seen On Screen. Now doing over 2 billion in turnover a year
- whatrocks 8y ago
  Not surprising at all! Instagram is probably next for this, then. Every single non-ad post could be a way to sell you the clothes or widgets that your friends are wearing/using in their fabulous lives.
esturk 8y ago
Better be opt-in or it would be annoying.

petard 8y ago

Working in this area the biggest challenge is getting the details right. Small difference like cut-outs on a dress make a huge difference in terms of end-user relevance. Same for colors, millennial orange comes in many shades. Deep Learning is pretty bad with such details.