( *via Hua Meng )
Recently, Alibaba Turing Lab ( Alibaba Turing Lab ) scored 74.8 points in the Object Detection Competition 3 : TRAIN ON PASCAL VOC DATA project of the Pascal VOC Challenge ( Pattern Analysis, Statical Modeling and Computational Learning ). , which refreshed the world's best result in this test. (Result query URL: http://host.robots.ox.ac.uk:8080/leaderboard/displaylb.php?cls=mean&challengeid=11&compid=3&submid=16634#KEY_ATLSSD )
It is reported that the PASCAL VOC Challenge is a benchmark for classification, recognition and detection of visual objects, providing standard image annotation datasets and standard evaluation systems for detection algorithms and learning performance. In the field of computational vision, the Pascal VOC challenge is similar to Goldbach's conjecture in mathematics, and is the world's top competition with ImageNet and COCO ( Common Objects in Context ).
PASCAL VOC's photo collection includes 20 categories: human, animal ( birds, cats, cows, dogs, horses, sheep ) , vehicles ( airplanes, bicycles, boats, buses, cars, motorcycles, trains ) , indoor ( bottles, chairs, dining table, potted plants, sofa, TV ) . The data set has good image quality and complete annotation, which is very suitable for testing algorithm performance. It is a "standard" data set for algorithm performance testing in the field of deep learning images.
Researchers from Alibaba Security Turing Lab said that PASCAL VOC has one of the most authoritative and general datasets in the world, and it is also extremely challenging. To this end, they developed the ATLSSD algorithm ( Alibaba Turing Lab Single Shot Multibox Detection ). The algorithm is based on the one-shot detection framework SSD , which takes into account model performance and efficiency, and adds inter-layer connections in the multi-scale feature extraction layer, similar to the FPN ( Feature Pyramid Network ) structure; RFB ( Receptive Field Block ) is added to the Extra Network . Structure, increase the feature layer's ability to represent targets of different scales and different shapes; during the training process, the model uses Hard Negative Mining to preheat the model, and then uses Focal Loss for finetune , so that under limited data, it can achieve better results through learning from scratch . training effect.
At present, the algorithm model of Turing Lab has been widely used in Alibaba's core business lines such as Taobao, Alipay, Youku, UC , etc., such as the intelligent prevention and control of prohibited content in goods, fake goods and videos, and has been implemented through Alibaba Cloud. The products and solutions on the platform (such as content security, real person authentication, etc.) provide high-quality AI services for enterprises and institutions.
"In order to make the algorithm closer to the business and solve more practical problems. We have launched the smart canteen project. The self-developed image algorithm for brushing dishes has realized the intelligent identification and payment and settlement of dishes, and recorded in the cloud with the help of IoT devices. The food composition of each meal of a user’s daily meal.” A researcher from the Alibaba Security Turing Lab said, “Our research in the field of artificial intelligence has provided a strong driving force for the intelligent development of the Ali economy, and it has also contributed to hundreds of millions of people. Users create the ultimate experience.”
It is understood that Turing Lab is Alibaba Security's top artificial intelligence R&D team, focusing on technology research and development in the fields of computer vision, natural language processing, machine learning and deep learning. It has applied for more than 50 patents, and has repeatedly It broke the world record in the international authoritative evaluation (in March this year , in the ICDAR2017 Competition on Multi-lingual scene text detection ( MLT ) natural scene multilingual text detection competition has set the world's best result) , most of its members are Ph.D. or master’s degree from Tsinghua University, Peking University, Zhejiang University, Chinese Academy of Sciences, etc.
Product Trial Channel: Content Security
Aliju Security (http://jaq.alibaba.com) is produced by Alibaba Security Department. It provides Internet business security solutions for enterprises and developers, covering dimensions such as mobile security, data risk control, content security, and real-person authentication. , and took the lead in the industry to propose "business-centric security" to empower the ecosystem and share with the industry the professional security capabilities that Alibaba Group has accumulated over the years.