AI Companies Flex Muscles at Winter Olympics; 10-Month-Old 3D Digital Creation Startup Raises $50M; Training Vision Transformers with 2040 Images

China’s AI news in the week of February 13, 2022

AI Tech Overview at 2022 Winter Olympics

The 2022 Beijing Winter Olympics kicked off with the opening ceremony on Friday, February 4. We hailed the ongoing transnational competitions, applauded the splendid athlete performances, inspired by sportsmanships, and joined the frenzy of the Olympics mascot Bing Dwen Dwen. Of course, we also witnessed a wide range of AI systems claimed to transform traditional broadcast or elevate training efficiency. These burgeoning innovations gave an unprecedented look at the confluence of sports tech and AI.

Sign language interpreter: Tencent launched an AI sign language interpreter to broadcast the Winter Olympics games for the deaf community. Built on multi-modal interaction, 3D digital avatar simulation, machine translation, speech recognition, and natural language understanding, the AI interpreter can recognize sign language with an accuracy of 90%.

AI to recreate athlete movements in 3D: Freestyle skiers can reach speeds of up to 16 meters per second, or 35 miles per hour while pulling off complex rotation moves. Baidu released an “AI+3D” system that can capture and analyze athletes’ motions and create a computer-simulated 3D model of skiers. It can also overlay 3D models of different skiers in the same picture, allowing different performances to be compared side-by-side.

Disinfection and cleaning robots: Robots that can disinfect, transport deliveries, and provide navigation assistance are pervasive in Beijing’s Winter Olympics Village. Beyond that, the functions of robots also extend to cooking meals and serving cocktails.

Multi-language translation: Speech tech giant iFlyTek provides iFLYTEK Jarvisen, an AI smart translator, to break the language barriers between over 2,871 athletes from 91 countries competing at the Winter Olympics, making lunchtime more engaging.

Taichi Graphics Raises $50 Million to Make 3D Creation Accessible

Taichi Graphics, a 10-month-old startup that aims to democratize 3D content creation, has closed Series A funding with a $50 million investment co-led by Source Code Capital, GGV Capital, BAI Capital, and Sequoia China. Founded in April 2021, Taichi Graphics is building open-source graphics infrastructures (compilers, graphics algorithms) and cloud platforms (HPC and cloud computing) to make graphics technologies readily accessible for applications like augmented reality, physical simulation, special visual effects in films and games.

The story began in 2019 when Dr. Yuanming Hu, the co-founder and CEO of Taichi Graphics, created Taichi. Taichi is essentially a Python-based programming language that can help developers to create 3D graphic renderings with significantly fewer lines of code and run on CPUs and GPUs much faster. By efficiently processing 3D visual data, Taichi promises to save a huge chunk of costs in large-scale 3D simulation, rendering, and vision tasks while maintaining comparable performance. For example, Dr. Hu’s team managed to recreate the 3D special effects in the movie FROZEN with 99 lines of code. The Taichi paper was accepted at the top computer graphics conference SIGGRAPH 2019. You can find its GitHub here: https://github.com/taichi-dev/taichi.

After Dr. Hu obtained his P.hD at MIT in March 2021, he returned to China and founded Taichi Graphics. This year, the company introduced Taitopia, which is claimed to be the world’s first cloud-native 3D digital content creation platform.

Taichi Graphics is one of the numerous startups that benefit from the expansion of MetaVerse, an interdisciplinary concept that builds on the foundation of cheap, scalable computing capacity, AI, and most importantly 3D simulation technology.

The company said the funds will be used to support the ongoing efforts in the development of the next-generation parallel programming language, the growth, and expansion of the Taichi community and ecosystem, as well as investments in our ground-breaking 3D digital content creation platform Taitopia.

Training Vision Transformers with 2040 Images

A team of researchers from China’s Nanjing University studied how to train Vision Transformers, the new rockstar in the computer vision community, with only 2040 images (why pick this number, just curious) while still obtaining state-of-the-art results.

The researchers proposed a method named Instance Discrimination with Multi-crop and CutMix (IDMM), a combination of tricks that are well-suited to train with small data, and investigated why these tricks are effective on small data. With IDMM, researchers achieved SOTA on 7 small datasets. They also discovered the transferring ability of small datasets is good when pre-trained on small datasets.

It’s quite an interesting paper as it proposes a new method against the fashion of training Transformers with hundreds of millions of data. Instead, it focuses on small data and explores the viability. You can read the paper here: https://arxiv.org/abs/2201.10728.

Investment News:

Rokae, a Chinese lightweight industrial robot supplier, has raised almost RMB400 million yuan ($63 million) in its Series C funding round. Founded in 2014, the Beijing-based company is committed to providing the industry with high-performance industrial robots and intelligent manufacturing solutions.

Feng1, a smart vending machine company, has bagged RMB300 million ($47.1 million) in its Series A funding round led by SoftBank. Founded in 2017, the Shenzhen-based company provides services in 28 Chinese cities with more than 50,000 terminals, serving 10 million users.

HoteamSoft, an enterprise software company, has closed its Series C fundraising with an RMB400 million yuan ($63 million) investment. Founded in 1993, the Shandong-based company provides software solutions for industries like aerospace, nuclear power, and rail transit, as well as machineries such as automobiles, construction tools, molds, bearings, and high-tech electronics.