The Computer Vision Center and the University of Barcelona have developed CLOTH3D, which is the first large-scale 3D synthetic data set that can perform virtual tests on clothing of different shapes.
With the current health crisis Coronavirus disease Internet shopping has become popular, and clothing cannot escape this trend. However, it is not easy to know how the garment feels to a particular body or know in advance whether the size is appropriate.
A group of researchers Human body posture restoration and behavior analysis Development of Computer Vision Center (CVC) and University of Barcelona (UB) CLOTH3D To solve this problem.The project includes the first large-scale comprehensive data set 3d human sequence Dressed, recently published in a magazine Computer Vision-ECCV 2020 Seminar.
“These data consist of thousands of sequences in which different people perform various actions, such as dancing, walking, jumping, climbing or posing. These human bodies are represented as 3D objects, and we use a random and random way The algorithm that generates clothing automatically wears them. Different materials (Cotton, denim, silk and leather). After that, we use a tool that physically simulates the fabric,” Sergio Escalera, the lead author of the CVC and UB study, told SINC.
The result is that the clothes present a realistic sequence of dynamics and deformations.thank you very much Process randomness From the production of clothing, scientists have obtained thousands of sequences, one of which is no longer repeated.This huge variability is for it in the so-called Deep learning.
In this sense, artificial intelligence and deep learning (Deep learning) Plays a key role in 3D clothing modeling and generation. Researchers believe that by simplifying the work of designers and animators, these models will greatly improve the experience of the virtual fitting room.
“Deep learning has recently shown great progress in the 3D field. However, these types of models require a lot of data to learn. In terms of clothing, we have not found enough public data sets to develop models based on deep learning. Because our data provides a wide variety of clothing and realistic physical characteristics, it is possible to improve or create new models that can produce more realistic predictions,” Escalera added.
So far, most of the models used to simulate clothing of different body shapes are 2D. This is because 3D models require a lot of data.
“The best person to take advantage of this type of technology is Virtual dressing roomThe scientist explained that it is therefore necessary to learn how to model hundreds of different garments with different materials, sizes and topologies. “
Get 3D data
The three main strategies for generating 3D data of people wearing clothes are: 3D scanning, 3D generation of conventional images, and synthetic generation. The researchers said that 3D scans are expensive and cannot distinguish the human body from the clothes. “That is to say, at most they can extract 3D shapes, as if the human body and clothes are a single object.”
On the other hand, infer 3D geometry They say that clothing in traditional images is “inaccurate and cannot adequately model the dynamics of clothing”. Finally, synthetic data is easy to generate and has no measurement errors.
“Because the development of 3D models requires a lot of data, we decided to generate our own data. Hugo Bertiche, the co-author of the book, added that we have designed and released the largest data set of this type, which includes a variety of clothing And all kinds of clothing sports.
CLOTH3D Clothing types, shapes, sizes, tightness and fabrics vary greatly. Thousands of clothing with different postures and body shapes can be simulated to create very realistic clothing dynamics. In terms of the variability of clothing, shape and posture, CLOTH3D is unique and includes more than 2 million 3D samples.
“We have developed a production line that can create a unique collection for each sequence based on clothing type, topology, shape, size, fit and fabric. Although other datasets contain very few clothing, our dataset has several Thousands, making it the largest data set in the field today. However, we did not participate in the development alone, but published it in an open-access way so that all types of public can access it,” Escalera said.
This random clothing generator covers most of the clothing we can see every day (T-shirts, shirts, tops, pants, skirts, dresses and jumpsuits), they have different cuts, sizes and shapes. However, the variability of clothing in the real world is only limited by imagination. Therefore, the work of supplementing these data with new clothing must be continued indefinitely.
“on the other hand, Clothing simulation technology The use of computer physics is constantly evolving. This allows the creation of new data with a higher level of detail (possibly down to the thread level) and reduces computational costs in a more efficient way. Finally, given that the purpose of this data is to help develop new models, designing new models is part of the next steps in this direction.
Other utilities of the project
However, the textile industry is not the only industry that can use this data set,” Entertainment industry “This may also benefit, because CGI movies and video games may be more realistic,” Bertiche believes.
“We have discovered the interest of many private companies. However, it is currently open to all types of public only when all types of the public intend to use it for scientific purposes.” Escalera said.
The research data is public so that researchers around the world can use it to design and train models based on deep learning to better reflect the behavior of 3D clothing. “in 3D clothing, The public data set is relatively sparse, and clothing changes are small. Therefore, our data plays an important role in this situation”, the expert concluded.
Development of test items. / CVC
ladder. “CLOTH3D: Humans in 3D clothing”. Computer Vision – ECCV 2020