Autograding (for Clothing)

for 15-869K, April 24th, 2020

Grading is the process of creating different sizes of a single pattern to fit different sizes of people.

(Image from AliExpress listing: https://www.aliexpress.com/item/4000351752304.html)
Which of these necklaces fit?
Does size x fit...
  • ...if x < r?
  • ...if x = r?
  • ...if x > r?
Hmm, this isn't a very useful constraint.

Fit is both practical and aesthetic.

A more useful formulation:
x = r + w + d
"size" = "body size" + "wearing ease" + "design ease"
Wearing ease is the space you add for practicality
Design ease is the space you add for for aesthetics
So, in our example:
wi = ?
di = ?
Say we wanted to manufacture necklaces.
We run tests to determine w. Our designers decide on d.
How should we pick which sizes [x1, ...] to make?
Use math! We'd like customers to find a "close to ideal fit":
minimize E[mini (xi - (1.1r + 5.5))2]
"minimize the average (over customers) squared difference between the ideal fit and the best available fit"

Problem 1: People aren't generally 1D.

need n1 points

need n2 points

need n3 points

(And being wrong in any dimension of fit is bad.)

Problem 2: People's measurements aren't strongly correlated.

So if your sizes cover 90% of people's wrist measurement, and 90% of people's waist measurement, and 90% of people's neck measurement...

...they only cover 0.9m of people (e.g., 72.9% for m = 3)

Standardized sizing will never be great, because people have high variance.

Research Question:
How do we automatically adapt a pattern ("grade" a pattern) for a given set of measurements?

See: http://www.inrialpes.fr/sed/people/boissieux/GARMENT_TRANSFER/paper.html
See also, (full talk): https://www.youtube.com/watch?v=4nNHZYsp9u4

Research Question:
If we resize a cut-and-sew pattern, how do we make sure fabric patterns remain consistent?

See: https://igl.ethz.ch/projects/aligned-seams/

Grading is essential, but difficult, and remains an area of active research.