Shenlan technology wins the double track champion of ECCV 2020 billion pixel image and video visual challenge

category:Internet
 Shenlan technology wins the double track champion of ECCV 2020 billion pixel image and video visual challenge


The gigavision 2020 challenge, which won the double crown of Shenlan technology, was organized by Tsinghua University Based on its newly launched data set panda, the worlds first billion pixel video data set. In the past decade, human centered computer vision analysis tasks, such as pedestrian detection, tracking, action recognition, anomaly detection and attribute recognition, have attracted great attention. In order to promote new algorithms to understand complex crowd activities and social behaviors in large-scale real-world scenes, billion level pixel target detection, which can enlarge the image by thousands of times, will be used in face recognition It has been widely used in many fields, such as mobile phones, unmanned driving, monitoring and security.

The task of gigavision2020 challenge is two classic visual tasks: image multi category target detection and video multi-target tracking. Task one is pedestrian and vehicle detection, and task two needs to extract the track of pedestrians in video. The accuracy and efficiency of detection on panda must be solved at the same time. The accuracy is challenged by obvious target scale changes and complex occlusion, while the efficiency is greatly affected by the resolution of 1 billion pixels. At the same time, the huge scale change of similar targets, the demand for long-term and long-distance tracking, and the complex scene of pedestrian congestion and mutual occlusion bring serious challenges. Deep blue AI team of Shenlan technology decouples the task into multiple subtasks in the detection track, and focuses on solving the difficulties independently. According to the previous detection experience, through the accumulated modules and methods, the results have been further improved. In the end, the difference between the second and third places was only a few thousandths, while the team of deepblueai was 2.6 percentage points ahead of the second place in the best performance, and won the championship strongly. As one of the core technologies of Shenlan science and technology, computer vision has been verified in a number of international top competitions, and has been applied to products in different fields, including Shenzhen AI thermal visual behavior monitoring system owl, which played a great role during the epidemic, and panda intelligence with automatic driving road test license plate has been obtained in Guangzhou, Shenzhen, Wuhan, Shanghai, Changsha and other places Buses, as well as all kinds of robots and intelligent retail products. The leading of basic research and algorithm will directly affect the performance of products and the application of artificial intelligence. Therefore, Shenlan technology will make unremitting efforts in the vision of artificial intelligence, serving the peoples livelihood. Source: editor in charge of mass news: Chen Tiqiang_ NB6485

The task of gigavision2020 challenge is two classic visual tasks: image multi category target detection and video multi-target tracking. Task one is pedestrian and vehicle detection, and task two needs to extract the track of pedestrians in video. The accuracy and efficiency of detection on panda must be solved at the same time. The accuracy is challenged by obvious target scale changes and complex occlusion, while the efficiency is greatly affected by the resolution of 1 billion pixels. At the same time, the huge scale change of similar targets, the demand for long-term and long-distance tracking, and the complex scene of pedestrian congestion and mutual occlusion bring serious challenges. Deep blue AI team of Shenlan technology decouples the task into multiple subtasks in the detection track, and focuses on solving the difficulties independently. According to the previous detection experience, through the accumulated modules and methods, the results have been further improved. In the end, the difference between the second and third places was only a few thousandths, while the team of deepblueai was 2.6 percentage points ahead of the second place in the best performance, and won the championship strongly.

As one of the core technologies of Shenlan science and technology, computer vision has been verified in a number of international top competitions, and has been applied to products in different fields, including Shenzhen AI thermal visual behavior monitoring system owl, which played a great role during the epidemic, and panda intelligence with automatic driving road test license plate has been obtained in Guangzhou, Shenzhen, Wuhan, Shanghai, Changsha and other places Buses, as well as all kinds of robots and intelligent retail products. The leading of basic research and algorithm will directly affect the performance of products and the application of artificial intelligence. Therefore, Shenlan technology will make unremitting efforts in the vision of artificial intelligence, serving the peoples livelihood.