Machine Learning can Help Customize Clothing Designs

A pair of meticulously handcrafted socks is the oldest known knitting item that dates back to Egypt in the Middle Ages.

Machine Learning can Help Customize Clothing Designs" />

Even though handcrafted clothes have occupied people’s closets for centuries, a new range of high-tech knitting machines has redefined how individuals’ favorite pieces are created.

However, such systems, which have produced anything from Nike shirts to Prada sweaters, cannot be said to be seamless. It can be a tedious and complicated process to program machines for designs—for example, if a person has to specify every single stitch, then even a single mistake can disfigure the entire garment.

Now, in a couple of recent studies, a research team from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) has developed a novel method to simplify the process—an innovative system and design tool for automating the knitted garments.

In one study, the researchers produced a system known as “InverseKnit” that converts images of the knitted patterns into instructions that are subsequently utilized with machines to make clothing. Such a technique can allow casual users to produce designs without any memory bank of coding knowledge. It can even allow them to resolve problems of waste and inefficiency that occur during manufacturing.

As far as machines and knitting go, this type of system could change accessibility for people looking to be the designers of their own items. We want to let casual users get access to machines without needed programming expertise, so they can reap the benefits of customization by making use of machine learning for design and manufacturing.

Alexandre Kaspar, Study Lead Author and PhD Student, CSAIL, MIT

The system has been described in the latest paper.

In another paper, the scientists developed a computer-aided design tool for modifying knitted items. The tool allows novices to use templates for altering shapes and patterns, for example, adding vertical stripes to a sock, or a triangular pattern to a beanie. One can image users making items that are tailored to their own bodies, while simultaneously personalizing for preferred aesthetics.

InverseKnit

It is a well-known fact that automation has already changed the fashion sector, with potential positive residuals of modifying the manufacturing footprint as well.

In order to get InverseKnit up and running, the researchers initially produced a dataset of knitting instructions, and also the corresponding images of those patterns. Next, they trained their deep neural network on that specific data to understand the 2D knitting instructions from images.

This may appear something like giving the system an image of a glove, and then allowing the model to create a series of instructions, where the machine subsequently obeys those commands to yield the design. When testing the InverseKnit, the researchers noted that it generated a precise set of instructions 94% of the time.

Current state-of-the-art computer vision techniques are data-hungry, and they need many examples to model the world effectively. With InverseKnit, the team collected an immense dataset of knit samples that, for the first time, enables modern computer vision techniques to be used to recognize and parse knitting patterns.

Jim McCann, Assistant Professor, Carnegie Mellon Robotics Institute

Although the system presently functions with a tiny sample size, the researchers are hoping to widen the sample pool to use InverseKnit on a larger scale. At present, they used only a particular type of acrylic yarn, but are hoping to test other different materials to render the system more flexible.

A Tool for Knitting

Although many developments have occurred in the field—like Carnegie Mellon’s automated knitting processes for 3D meshes—these techniques can be usually ambiguous and complicated. 3D shapes have inherent distortions that limit one's understanding of the items’ positions, and this can pose a burden to the designers.

In order to deal with this design problem, Kaspar and his coworkers created a tool known as “CADKnit,” which utilizes photo editing techniques, CAD software, and 2D images to enable casual users to modify templates for their knitted designs.

Through this tool, users can design both shapes and patterns in the same interface. In the case of other software systems, one might lose some amount of work on either end when modifying both.

“Whether it’s for the everyday user who wants to mimic a friend’s beanie hat, or a subset of the public who might benefit from using this tool in a manufacturing setting, we’re aiming to make the process more accessible for personal customization,” stated Kaspar.

To test the usability of CADKnit, the researchers allowed novice users to produce patterns for their garments and tune the shape and size. The users reported in post-test surveys that they found it easy to control and modify their beanies or socks and successfully created numerous knitted samples. They also said that they found it difficult to design the lace patterns correctly and would gain from a rapid realistic simulation.

Conversely, the system is just an initial step towards complete garment customization. The researchers discovered that clothes with complex interfaces between varied parts—like sweaters—did not work suitably with the design tool. The trunk of sleeves and sweaters can be joined in numerous ways, and the software is yet to have a way of describing the entire design space for that.

Moreover, the new system can use only a single yarn for a shape; however, the researchers are hoping to enhance this by adding a stack of yarn to each stitch. To facilitate work with larger shapes and more intricate patterns, the team is planning to utilize the hierarchical data structures that do not integrate all stitches, but only the required ones.

The impact of 3D knitting has the potential to be even bigger than that of 3D printing. Right now, design tools are holding the technology back, which is why this research is so important to the future.

Jim McCann, Assistant Professor, Carnegie Mellon Robotics Institute

The InverseKnit paper was presented by Kaspar alongside MIT postdocs Petr Kellnhofer and Tae-Hyun Oh, MIT undergraduate Jacqueline Aslarus, PhD student Liane Makatura, and MIT Professor Wojciech Matusik. The paper was presented at the International Conference on Machine Learning in June 2019, in Long Beach, California.

A paper on the design tool was headed by Kaspar along with Matusik and Makatura.

(Video credit: MIT CSAIL)