A survey of 1000 online shoppers by British consumer analysis firm Savvy Marketing in 2016 showed that 63% of buyers returned women’s clothing. It was found that online buying and returns were more in the 18-24 age group. Students in this age group went on to say that they bought a lot of clothes because they did not want to be seen with the same clothes on Facebook everyday. They also admitted ordering different sizes because it was impossible to know which would fit. As a result, they ended up sending back half of what they bought.
Many of us believe that shape is critical to fashion. But those avid watchers of the history of retail garments would know that the answer was actually a resounding NO. Shape was not important for most mass merchandisers. They who defined consumer street fashion accommodated 3 billion plus women, and an equal number of men in around a dozen sizes. In India, it is less than half a dozen for most brands - XXL, XL, L, M and S. And every one who wanted branded clothing but could not afford customised haute couture, would get themselves lumped in sizes that actually did not fit them.
Worse, each brand had a different measurement for every size. So you could be L for Marks and Spencer and still be XL for Calvin Klein’s. So consumers buying ‘ready mades’ were a confused lot. Especially the millennials who believed that going shopping to a mall was a drag. It was far more satisfying for them to browse and buy where the choice was immensely more -- and you could return at will. All that without having to visit the stores even once. The returns in the fashion segment alone accounts for a quarter of the total returns of online sales. Women’s fashion has the most returns and sixty percent of the returns are due to size. One in four loose fitting dresses like T shirts and loungewear are returned by online shoppers due to size misfit while for body hugging apparel like jeans and tops the returns are as high as 50%.
Garment majors across the world have been grappling with the problem of standardising the neck size and chest size of consumers. Usually, sizes differ every two inches except between size 36 to 40 where the clothing could be with variations for every inch. Now a New Jersey based startup, founded by CA and businessman Arup Chakraborty, along with two professors from IIT Delhi, have created an AI based solution for sizing with accuracy up to half an inch. The company which has set up a research centre at a South Delhi neighbourhood with a dozen engineers is bootstrapping and creating a global solution at a low budget with high end programming skills of Indian techies. The AI solution will help merchandisers and retailers attain garment sizing accuracy below half an inch (around 1 cm) which is unprecedented. The technology is based on the development of a model mesh based on two photos of the individual from an ordinary mobile phone of any consumer. All you have to do is keep the phone straight and upload a front side image and a sideways image onto their app.
It is a challenge to develop a three dimensional shape from a front and a side posture photo of any individual. “How would you know in a simple binary image that these are my hand points not all in the same plane? To find that out we did anthropometric identification.” says Arup Chakraborty. “Human bone structure is pretty predefined and catalogued. So we trained thousands of body sets with bones to get the points accurately. Once we got those defining points, it gave us a perfected linear image. The next problem to solve was defining the circumferential, measure with accuracy since there was no depth”.
Micheal Black is a much awarded global expert on computer vision based in Germany. He uses advanced machine learning and graphics to solve problems in the clothing industry and is called the guru of the mesh. In 2013, he cofounded Body Labs Inc. a startup that developed advanced machine vision and optical flow estimation for 3D human shape and motion analysis. He researched with a team of 50 computer vision experts in New York for two years to conceptualise the mesh out of 2D binary images. The company perfected estimation of body shape to the accuracy of four inches ( 10 cms) as per a paper published by Professor Black. Body Shape was bought by Amazon in 2017 for an undisclosed sum, that was estimated to be between $50 to $70 million as reported by Tech Crunch. His work is no longer available in the internet as he is under a standard non disclosure agreement NDA. Amazon has also reportedly taken a patent on the back end processing of Professor Black’s work.
So to discover how artificial intelligence converts two 2D images into a 3 D image with accuracy, this author travelled to IIT Delhi to meet Professor Sudipto Mukherjee, a renowned computer scientist who is also the Dean Faculty. “Arup came to us in August 2017 with an interesting problem to solve. We initially thought it was an ordinary computer science problem where you had an image and you need a measurement. The conventional way is to apply geometry from images. But when we started to work that way, we found it cannot be done due to the limits of the speed of activity and the average phone resource with 1000 x1000 pixel resolution” says Prof. Mukherjee. So the work was split into multiple processes and each process was solved differently.
The team combined the deep learning of artificial intelligence which is very good at finding the gross profile with conventional classical solutions that are great for defining small areas. The entire work was broken down into around 16 to 17 modules. While Professor Mukherjee wrote the algorithms for each module at IIT Delhi, like say ‘for creating an edge’, or for ‘creating a centre line’ or for ‘finding the shortest path between a point and a curve’, the programming and the extensive backend coding was done mostly at the basement office by Prof. Chakraborti and his young band of engineers. Some of the coding work initially was also done at IIT Delhi. It was an amalgamation of each of the processes, some solved by classical methods and some by AI that were cascaded to create the total solution.
As a first step, the binary images taken by the consumer were extracted with great accuracy from the background. But they did not synthesise the measurements directly from the 2D image of the human body. They instead added a model placed on the mesh. Then they used the image of the human body and the image of the model and tried to match the two images. They worked back and forth from the human image to the model image and started stretching and pulling the model to get the right measurement and the final shape in the background. It has taken the startup and the professors around eight months of back breaking work and thousands of hours of coding to improve and optimise the accuracy, speed and efficiency of the programme that brings AI and classical programming together to create a much needed solution for the garment industry.
Will this be one of the first success stories of Made in India in the garment industry commercial applications from the field of Artificial Intelligence? It remains to be seen.
The views expressed above belong to the author(s). ORF research and analyses now available on Telegram! Click here to access our curated content — blogs, longforms and interviews.
Sandip Sen is an author and journalist writing on a vast range of subjects from economy to technology environment to lifestyle. He is a regular ...
Read More +
PREV
