Shulou(Shulou.com)06/02 Report--

2019-09-18 17:19:34

Unwittingly, AI has penetrated into all aspects of life, including what we know as food takeout.

From the menu entry of the merchant, the identification of the signboard, to the selection of the first image of the advertisement automatically generated by AI during the promotion, and the identity verification of the takeout rider. Meituan's AI vision ability has penetrated into all aspects of his business.

Wei Xiaoming, head of Image and Video Group of Meituan Visual Image Center of ▲

Recently, Zhidong came to Meituan headquarters in Beijing and had an in-depth dialogue with Wei Xiaoming, head of the image and video group of Meituan Visual Image Center, and comprehensively interpreted Meituan's visual AI capabilities, Meituan's visual AI platform development process, and the "power plant" behind Meituan Visual AI.

First, from menu input to AI selection, AI is everywhere

'unlike many other companies, Meituan's AI technology has strong business-oriented features, 'Mr. Wei said.

At present, Meituan's AI technology is mainly divided into four categories, namely: voice / semantic understanding technology based on AI, visual processing technology based on AI, distribution scheduling optimization based on operations research, and unmanned distribution technology based on autopilot.

▲ Meituan AI visual layout

Wei Xiaoming is the head of the image and video group of Meituan Visual Image Center. He has more than 9 years of experience in visual research and technology management. He has previously worked at Canon Research Institute and Samsung Research Institute.

Since joining Meituan in 2015, Wei Xiaoming has led more than 50 AI vision projects. In the interview, Wei Xiaoming introduced Meituan's typical AI vision application scene from the perspectives of merchants, riders, users and platform.

▲ AI menu photo entry

For merchants, Meituan AI enables them to input menus through photos, extract structured information from paper menus through text detection, semantic segmentation, and visual relationship learning, and reduce the entry time of merchants' menus from small hours to less than one minute.

▲ AI signboard recognition

With the continuous development of OCR technology, menu recognition, signboard recognition and other subdivision scenes have become the current research hotspot in the field of OCR. At this year's ICDAR 2019, Meituan hosted the industry's first Chinese face signboard character recognition competition (ICDAR 2019 Robust Reading Challenge on Reading Chinese Text on Signboards) and released the industry's first real scene signboard image data set.

In addition, behind the merchant's certificate qualification authentication, the rider's facial scan authentication, the user's AI dish recognition Mini Program, the selection of the first image of the advertisement, the rider's spot check, and other applications, there is Meituan AI visual blessing.

Wei Xiaoming said that at present, Meituan has tens of thousands of technical personnel, Meituan AI Vision Center also has dozens of employees, and the team size is still expanding.

Second, the GPU computing platform has been upgraded in an all-round way, resulting in a hundredfold increase in efficiency

Meituan's AI Vision team was established in 2015 to provide Meituan with AI vision capabilities (such as image review, intelligent image selection, etc.). The period from 2015 to 2016 can be regarded as the first stage of the development of Meituan's AI vision platform.

With the soaring demand for AI computing power within the group, Meituan carried out a comprehensive upgrade of the enterprise-level computing platform in 2017, upgrading from a CPU-based computing platform to a clustered AI computing platform based on Nvidia GPU. 2017-2018 can be regarded as the second stage of the development of Meituan AI vision platform.

After upgrading to a clustered AI computing platform based on Nvidia Tesla V100 GPU, Meituan AI can achieve a hundredfold improvement in offline training (Training) in text detection, face recognition and commodity recognition.

In terms of applied reasoning (Inference), Meituan's collocation based on Nvidia Tesla P4 GPU + TensorRT can improve the computational efficiency by tens of times. Moreover, Meituan currently uses FP32 precision, and the performance will be further improved if Nvidia T4 Tensor Core GPU is used for FP16 precision reasoning calculation in the future. This low-latency and high real-time computing performance is very important for the experience of users and dispatchers.

For example, in order to ensure the safety of users, the current 700000 riders of Meituan Daily have fully covered the "facial scan authentication" function. But this feature adds an additional verification process for riders, which has an impact on efficiency and experience.

Therefore, in order to ensure the efficiency of face comparison of riders, Meituan uses the scheme of GPU parallelism + TensorRT for large-scale face comparison of riders. Compared with the CPU-based scheme, this scheme can be accelerated by more than 20 times, and the speed increase will make the delivery boy's "facial scan authentication" process faster and better.

Wei Xiaoming said that at present, the average daily invocation of the AI vision service of Meituan server cluster has reached hundreds of millions of times.

The next step of Meituan AI

In 2018, combined with Meituan's computing platform upgrade, Meituan AI Vision got a very large-scale landing, covering the above-mentioned menu recognition, face authentication, face scanning payment, advertising generation and many other intra-group application scenarios.

▲ Meituan won the second place in CVPR2019-FGVC6 Commodity Identification Competition.

Since 2019, Meituan's AI Vision team has not only supported the AI needs within the group, but also gradually made a voice in large-scale international competitions. In 2019, Meituan AI Vision team won the Top3 position in CVPR, ICME and other famous visual competitions.

Wei Xiaoming believes that the current AI algorithm is still in the process of fast iteration. For a large platform like Meituan, continuous iteration of TensorFlow, Caffe, MXNet and other deep learning frameworks can improve the efficiency of computational parallelization, which is very important for the optimization of Meituan's specific AI scene.

In the next step, Meituan AI Vision team will also expand the landing of more AI scenes such as video understanding, store digitization, unmanned distribution system and so on.

Conclusion: the application of AI breaks out in an all-round way, and the requirement of real-time performance is constantly improving.

With the outbreak of AI applications, many enterprises are upgrading the AI computing platform, especially the AI reasoning applications with high real-time requirements, such as facial scanning authentication and photo information retrieval, the traditional computing platform has been unable to meet the needs.

Unconsciously, AI has gradually penetrated into all aspects of our lives, in you inadvertently, may have enjoyed the convenience of life brought by AI.

Https://www.toutiao.com/a6737937713061691908/

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.