Recently, the first WIDERFaceandPedestrianChallenge2018 (hereinafter referred to as WIDERChallenge) jointly organized by SenseTime, the Chinese University of Hong Kong, Amazon, Nanyang Technological University, and the University of Sydney, came to a successful conclusion. As a new global top computer vision competition, the challenge attracted more than 400 teams from all over the world to sign up. In the whole process of competition organization, design, data selection, etc., SenseTime has given full play to the advantages of industry-university-research collaboration, and put forward many competition topics and rules that meet the needs of practical scenarios and industrial applications, and promote the industry with brand-new industry norms and standards. Exchanges with academia open up new development directions for computer vision research.
Develop a new standard data set to promote the emergence of high-performance algorithms
With the increase in application requirements and scenes, face detection, pedestrian detection, and person detection have become popular projects in computer vision research. The challenge focused on these three hotspots, designed three sub-tasks of WIDERFace, WIDERPedestrian and WIDERPersonSearch. At the same time, based on the ever-increasing demand for complex scenes, it enabled a data set closer to the real scene to enhance the practicality, innovation and challenge of the competition. Further promote the emergence of high-performance algorithms for face and pedestrian detection in the field of computer vision.
The WIDERFace face detection dataset is a standard dataset in the field of face detection. The WIDERFace dataset (http://mmlab.ie.cuhk.edu.hk/projects/WIDERFace/) was collected by Shangtang-Chinese University of Hong Kong Joint Laboratory in 2016, annotated and published as an oral report in the top computer vision of the year Conference CVPR conference. In more than a year, WIDERFace has become a standard data set widely used in the field of face detection. Compared with the previous face detection data set, the WIDERFace data set has an order of magnitude improvement in data difficulty, the number of pictures and annotations.
WIDERPedestrian provides a large-scale data set dedicated to pedestrian detection. Considering the two popular applications of pedestrian detection (surveillance and autonomous driving), the selection of pictures collected by surveillance cameras and on-board cameras has very different shooting angles, pedestrian scales, and illumination. Participants must propose more robust and universal Sexual methods to deal with different scenarios. And part of the data comes from night scene shooting, which brings greater difficulty to detection. At the same time, compared to other pedestrian detection data sets, WIDERPedestrian provides two different pedestrian annotations, pedestrians and cyclists, with more detection targets and pedestrians of different scales. The density of detection targets, the smaller pedestrians and the occlusion between pedestrians, etc. have brought more challenges to the contestants.
WIDERPersonSearch is a novel task of retrieving characters from 192 movies. It needs to retrieve all instances of corresponding characters from a large database based on the standard photos of actors. Character retrieval is not only practical, but also very challenging. The standard photos of actors and their costumes in movies are often completely different. Even in the same movie, the actor's clothing and environment can change dramatically. In addition, there are various obstacles that increase the difficulty of recognition, such as dark light, occlusion, and blurred motion. This more challenging task provides a new stage for the future development of character search algorithms.
More than 400 teams from all over the world have participated in the current three stand out
The WIDERChallenge challenge attracted 432 people/teams from academia and industry around the world to register, and 73 teams submitted their results. Participants come from China, Russia, Japan, the United States, Australia and many other countries. The participating institutions include the Institute of Computing Technology of the Chinese Academy of Sciences, Microsoft Asia Research Institute, Peking University, Shanghai Jiaotong University, University of Chinese Academy of Sciences, University of Science and Technology of China, NtechLab, Carnegie Mellon University, University of Hong Kong, Hong Kong Polytechnic University, University of Technology Sydney, Central Japan University, Santa Clara University, Georgetown University, University of Illinois, Munich University of Technology and other universities and research institutions, as well as commercial companies such as JD.com, Yahoo, Megvii, HKUST iFLYTEK, and Didi.
Faced with three extremely challenging and practically valuable sub-tasks, the participating teams showed their abilities. The three winning teams of the WIDERFace face detection task competition all used deep learning technology to design and implement face detection algorithms, and all used or borrowed the idea of ​​feature pyramid to enhance the features of the backbone network, and the initial matching template ( anchorbox) has been grouped and redesigned. The champion team also used model fusion to achieve better results.
In the pedestrian detection task under WIDERPedestrian monitoring and autonomous driving, the winner uses the traditional FasterRCNN method, using the ResNet network plus the pyramid structure (FPN) to extract and integrate multi-level semantic features. And by adding a cascade network (CascadeR-CNN) to the detection module to train a better boundingbox regression, using RoI-Align instead of RoI-Pooling in FasterRCNN to help detect small-scale pedestrians and use attention The mechanism (channel-wise attention) handles the occlusion problem.
In the third subtask WIDERPersonSearch, the winning teams split the task into two stages. The first stage is for face recognition, adding very similar faces to the query set (queryexpand). The second stage is to re-identify pedestrians, and use physical features to process images to be queried that cannot be accurately judged by facial features. Finally, the feature similarity between the face and the human body is integrated to obtain the sorting result. Both the champion and runner-up used Jacquard distance and Euclidean distance for sorting.
As the co-sponsor of the challenge, SenseTime did not send a team to participate in the competition. The award ceremony of the challenge will be held in Munich, Germany during the ECCV2018 conference in September this year, and related seminars will also be held. In addition, all the winners will be invited to co-write competition papers and make presentations at the ECCV2018 seminar. After the competition is over, the organizer will still open the test server of the verification set for participants to conduct scientific research.
SenseTime adheres to the mission of "Insist on originality and let AI lead human progress", and has been focusing on promoting the development of computer vision and deep learning technology. It not only has the original deep learning platform Parrots independently developed by itself and the world's top supercomputer center, it is Asia The largest AI research and development base, at the same time, with the mode of industry-university-research collaborative innovation, and with the rich application experience in face recognition, image recognition and other technical fields, it continuously promotes the close connection between academic and industrial circles. By holding the WIDERChallenge challenge, SenseTime has backed up academics with years of research accumulation and landing experience, formulated new industry standards, led the trend of industrial and academic development, and promoted technological progress in the field of computer vision.
Welding diodes are available for medium frequency (over 2KHz) and high frequency (over 10khz) applications.They have very low on - state voltage and thermal resistance. Welding diodes are designed for medium and high frequency welding equipment and optimized for high current rectifiers. The on-state voltage is very low and the output current is high.
Welding Diode,Disc Welding Diode,3000 Amp Welding Diode,High Current Rectifier Welding Diode
YANGZHOU POSITIONING TECH CO., LTD. , https://www.cnfudatech.com