Post-Mortem for Bra Theory

I remember sharing it with my partner who is in a less common cohort and had done plenty of research prior to my sharing this with them... they were not at all impressed and felt like there was something lacking in their methodology and priors.

Specifically, my partner has found that the most consistent and predictable sizings are from the UK -- everything else is more all over the place and must be tested on a per brand and pattern basis.

The post by A Bra Theory states it can take as many as 9 attempts with their system to achieve a fit.

With US sizes and her proper measurements, it can take 3-4 attempts for my partner to find something that fits or determine that they do not carry anything that will. With UK sizes and her proper measurements, there is typically success within 1-2 iterations.

One can find more information about this on the subreddit wiki for A Bra that Fits: https://www.reddit.com/r/abrathatfits/wiki/beginners_guide

I find it strange that in the past nor in the present the Bra Theory hasn't seemed to lean on A Bra That Fits or the UK sizings and patterns.

Fabric is a very hard problem because it has an extremely high number of degrees of freedom. Flexible materials are like this consistently. If you want automation, with few exceptions, you want stiffness.

Tesla has been unable to automate wire harness installation because it's super flexible with many effective degrees of freedom (they hope to solve that using a stiffer flex cable that consolidates the wiring). Even moving some fiberglass fluff with a robot was an unreliable bottleneck they eventually removed.

I even saw a project once that attempted to automate cloth handling by first stiffening the cloth with starch so it could be more predictably moved from place to place.

Cloth is fundamentally hard. It's not impossible to solve these problems, but it's not at all trivial.

Elsewhere on HN today:

https://news.ycombinator.com/item?id=21876063

> Uniqlo heads towards full warehouse automation with groundbreaking robot that can fold and box clothes

It is more than just the age of tailoring vs undergarment making that is the problem. If tailoring was the only issue, we would just tailor women's undergarments and declare success. The real issue is that women's bodies are far more variable than men's bodies. Most garment designing software gets around this by fitting exactly the hardest part of the body and then scaling everything else. Bras are one of the more difficult parts of the body to fit although the hip area (pants) are also difficult to fit for the female form.

Pants actually illustrates the problem in a way that everyone (men as well as women) would understand. You can't just take a few measurements and expect success. You have to know how the measurement is distributed. Is the added inches on the stomach? Is it on the hips? Is it on the posterior? I won't go into the many crotch measurements - but they are also important.

I buy my dress shirts at Indochino. They do multiple first measurements on you the first time, then they make a first iteration shirt. You come for a second fitting with that shirt when they look how it sits and re-adjust the measurements which are stored in your profile and all subsequent shirts are made to them. The result is pretty good. Never been to a custom tailor but I guess this is a good industrialized approximation of what they do. I am wondering if something like this could have worked.

Ever since I came across this video https://youtu.be/gy5g33S0Gzo?t=4 of a robot taking 1,5 hours to fold five towels, with mixed results, I think about how much of our brain might be devoted to handling fabric every time I have a towel in my hands. Probably the same with handling fluids, like predicting how milk is going to behave in an opaque carton.

> As a naive guy

I intend this reply to supplement your list of challenges, all of which are valid. I hope this helps convey an additional depth of complications in designing women's elastic support wear that may not be immediately apparent.

> You've got a data capture problem

The missing piece here is the intersection of: clothing "fit" problem, where each person's body will interact with clothing at different points; the gravity "g-forces" problem, where the clothing is required to bear weight at every step without tearing apart; and the appearance "fashion" problem, where different shapes carry that weight at different points.

I'll take apart t-shirts first, and then switch back to bras.

Given several men of slightly different torso composition, a simple men's t-shirt can interact with their belly, lower waist (shirt length), ribcage, nipples (ask any runner), armpit, upper arms, neck. At each of those interaction points, the clothing must fit 'correctly' as perceived by the human being. Each human being has a preference for each of the listed areas, and will complain about poorly-fitting t-shirts if they do not fit 'correctly'.

I require loose armpit and long length shirts, among those seven options, and have more or less given up on ever finding a solution for my abnormally-shaped ribcage. So I typically end up in men's XL, because it provides enough 'airspace' to smooth out my torso and extends low enough to keep me from flashing people.

Someone else with the same torso length, torso diameter, and waist measurements might choose a Medium instead, because they have a decked-out bod and want to show off every curve. They'll need shirts made with more stretch than mine, or else they'll rip the armpits open waving hello to a friend. Shirts designed for a tight fit use a different 'cylinder' cut than shirts designed for a loose fit, so you can't just put them in my sized-down XL because I've been selecting for a different cut of shirt than is appropriate for their desires — and you can't just shrink all dimensions on an XL by 30% to get a M because human body parts don't shrink at the same percentage rate.

So, you can evaluate the 'fit' of t-shirts based on these criteria, none of which can readily be captured by 'shape data' alone:

* Does the shirt show off, or mute/hide, your physique to the degree you desire?

* Does the shirt rub uncomfortably, or rip/tear, when you move while wearing it?

* Does the shirt fabric irritate your skin, at your desired tightness?

You can custom-bespoke every shirt to a person's shape, but you'll still have to make fashion decisions ('drape', for example) that will include some and exclude others. If you end up deciding to print custom-fit "perfect mirror of your shape" fashion, that is fashion, too, and I would loathe it with all my heart because that is not what I want my clothing to express.

Moving on from the t-shirt analogy to bras, there are additional problems that bras have to solve for, that t-shirts do not:

* Bras are held with elastic against an extremely sensitive area of skin for hours at a time (sweat is a primary concern)

* Bras need to provide front and back appearances that matches the desires of the wearer (padded, unlined; demi, full coverage, balconette, push-up; longline, racerback, strapless)

* Bras need to provide support for 'weight' anchored to both sides of the chest (US average ~3lbs/side, K cup ~8lbs/side), that can cause pain every time you take a step, encounter small vertical G-forces (stairs, elevators, cars), or large all-directional G-forces (cars, subways, sports)

* Bra elastic loses stretch over time, due to wear and tear from the thousands of G-force events per day they intercept and reduce bodily impact of (adjustable straps, multiple rows of hooks, discard and replace occasionally)

So, not only do you need shape data, you also need "fashion" data, "fit" data, "fabric" data, and "gravity" data. You need data about composition — density for compressibility, shape for support and fashion, total weight for structural integrity — that you can't measure at home easily if even at all. You need material that can stand up to being punched from within a million times that also feels comfortable when held skin-tight all day. You need to make it fashionable, while keeping it fashionable across multiple size vectors (band size can vary from 24" to 48" or more, cup size can vary from 1/AA to 8/H or more). You need to transfer G-forces from the wearer's chest to the elastic while not digging into their skin more than they can bear (bralettes, underwires). You need to plan for elastic weakening over time (adjustable straps, multiple sets of back hooks).

Many men only see that level of tailoring in bespoke suits/tuxedos, and wear them once a year or less, and suffer no consequences the rest of the year. Many women have to wear one every single day, or else they suffer chest pain (F=m*a with only skin and ligaments to bear it) and/or societal outrage. It's a really intense market to try and serve.

I hope this helps.

There is also the 2D to 3D problem, if you’ve ever tried modeling clothing on a mannequin to make patterns, you’ll notice how much skill it takes on a doll with perfect measurements. I think the problem here is there is way to much variance in shapes to come to a custom 1:1 for everyone. People are asymmetrical, and proportions vary wildly. I mean you could cut corners, but you probably end up generalize to the extend that it isn’t really custom fitting anymore if you want to scale this up.

I think maybe the problem is that they swung for the fences out of the gate — a general solution to automatically creating custom fitted traditionally made bras for any body type.

They could have maybe had more success by being less ambitious:

Focus on one body type that has a relatively easy solution to make the ‘perfect bra’

Make custom bras that are close and require a single trip to a tailor for small adjustments — or one that enables the customer to make markings somehow and send it back to the company to adjust.

A “smart bra” for fitting that has sensors that send measurements about stress and weight distribution.

Try and figure out new materials and construction methods that cover up for inadequacies in the fitting algorithm. Make something that looks unusual but works — like crocs shoes. Some people will hate it but others will swear by it.

There is a saying: "if you don't know how to do it, you don't know how to do it with a computer".

Yes, that would be really cool. I found some bras that fit me well for a time, but seem to wear out. By day, I'm a composites engineer, so I would love to use some really good patterns and see if I can make one with Kevlar or something that will wear out more slowly.

Reading the description, I though that there is enough interesting, high-level, material for a few phD thesis (algorithm, material science, industrialisation of a process, etc).

I would love it if the author found a way to keep the research effort going in an academic setting.

I agree with your take on this. Facebook has done considerable work with their Detectron / Dense Pose projects which can UV unwrap human bodies using AI. (https://github.com/facebookresearch/DensePose) Another is the smpl / smplify-x project (https://github.com/vchoutas/smplify-x)

Further, by utilizing generated (synthetic) data with 3d scanned human models, or generating your own with something like makehuman or daz3D, you should be able to find a way to UV unwrap a distorted sphere. I've had a lot of success with generating synthetic data like this with Blender 3D and training various neural networks (Mask RCNN / Deeplab for example) with various combinations of this synthetic and real datasets.

Yup, they have the wrong tech approach. Good they stopped wasting money though…

Back in 2009, HN also discussed "The Physics of Bras" (2005):

https://news.ycombinator.com/item?id=925236

There definitely are standard sizes, but there is enough variation between manufacturers and between wearers that comfort can't be achieved at the level all can agree on. My wife has difficulty in finding the right bra and I have difficulty in getting the right shoe ( which you also would expect to be standardised)

Clothing tends to be cut from patterns of 2D shapes that are then sewn together. In theory if you measure certain key parameters of the customer you can adjust a standard pattern to fit them without further refinement. Imagine if you took a thin man, fat man and normal man then measured them all, and designed a pattern based on the normal man. Then found weighting’s to scale the pattern to the thin man and fat man. You could then (in a naive fashion) interpolate between these weighting’s to create patterns for every man inbetween. That would be your algorithm.

They were trying to design custom-fit bras automatically, based on 10 measurements of the body. But 10 measurements are too few to capture the wide variety of boobs that exist, and it was too expensive to gather ground truth examples (because you have to pay for expensive custom fit bras for a load of women) so they failed.

Actually, having multiple views/angles and just capturing how the skin moves in 3d-space when forces are induced (through e.g. arm raising, jumping, or various other methods) should allow for inferring some elastic properties. This could allow detailed fitting with parametric stiffness/support trade offs.