Yi+ Breaks the World Record in Pascal VOC with Yi+'s Object Detection Accuracy Ranking at the Top
BEIJING, Sept. 12, 2018 /PRNewswire/ -- In July 2018, Pascal VOC comp4 object detection challenge, an internationally established computer vision competition, took place. The AI team of the Chinese company Yi+ took the first place in the Single Model for Object Detection Challenge, surpassing other famous companies at home and abroad. Moreover, the Yi+ AI team broke the world record with an accuracy of 90.7%, becoming the world's first computer vision enterprise with a total score exceeding 90%. At the same time, in Pascal VOC "comp3", the team broke through 80% of accuracy for the first time and set a new world record.
The results can be seen here: http://host.robots.ox.ac.uk:8080/leaderboard/displaylb.php?challengeid=11&compid=4
Pascal VOC comp4 is the exact place where many domestic and foreign enterprises' object detection teams prove their capability. The Yi+ AI team achieved accuracy of 90.7%, 1.5 percentage points higher than the second-place team. In the field of object detection, 1.5 percentage points of accuracy means that over 6,000 more bounding boxes needed to be detected. It is also worth noting that the AI team of Yi+ uses a more difficult Single Model, while other teams used a Multi-Model Fusion.
Pascal VOC is the initiator of visual recognition competitions which covers tasks such as object classification, object detection, and image segmentation. Pascal VOC has a profound impact on the development of computer vision, which can be felt from the task settings of the ImageNet challenge which basically followed the settings of Pascal VOC. Previously, Microsoft, Intel, CMU, Facebook, UC Berkeley and other top international R&D teams have all set new records in Pascal VOC challenges.
The Yi+ AI team leader mentioned that the method they applied in the Pascal VOC comp4 challenge is called FXRCNN, in which "X" stands for "diversity", meaning that the structure they used is more than just the high-accuracy model. Rather, it is applicable to multiple scenarios, including transporting to the mobile terminals (by changing Backbone and Head), implementing image segmentation (by increasing the Mask branch), realizing the detection of human pose key points (by increasing the Key point branch), and face detection (by increasing the Face branch).
The deep learning model submitted by Yi+, though having adopted the basic structure of Faster RCNN, has its own unique features: 1) using ResNeXt as the basic network, combining FPN to achieve multi-scale feature fusion; 2) in the post-processing stage, using SoftNMS and multi-Box electoral fusion; 3) pre-training through massive data of Yi+; 4) multi-scale training for further data enhancements. Through years of efforts in accumulating technology, solving practical problems, and continuously optimizing the internal model, Yi+ managed to realize a full-fledged improvement of the FXRCNN in dimensions such as speed, memory, precision and application range.
In the field of AI, object detection technology is widely used, and it is one of the key technologies for the Yi+ AI team. Yi+ applies object detection technology to image search engine and image and video structural engine. Yi+'s image search engine can detect objects from over 100 categories, covering apparel, 3C supermarket (Computer, Communication, Consumer Electronics), household products, daily necessities, transportation, etc.; Yi+'s image and video structural engine adopts the industry's most advanced deep-learning based general object detection algorithm, which supports object detection and recognition of more than 300 types of common objects, enabling detection, recognition, segmentation and tracking of scenes, people, vehicles, and objects (contours) in videos and images and can identify nearly 10,000 items and 400 scenes.
At present, Yi+ AI has applied object detection technology to multiple scenarios, including "smart hardware", "marketing", "new retail", "smart city" and so on. It is specifically applied to advertising platforms, new retail platforms, large screen AI assistants, smart security, smart transportation, smart communities and many other fields.
In recent years, Yi+ has developed and applied a large number of new technologies in the field of AI and big data. In addition to object detection technology, Yi+ has also achieved remarkable results in developing facial recognition technology. In March 2018, in the latest published test results of LFW (Labeled Faces in the Wild), Yi+'s facial recognition technology ranked the top worldwide with 99.83% accuracy and fairly low fluctuation range.
Facial recognition is one of the core products of Yi+ AI. Yi+'s facial recognition technology can quickly and accurately complete face detection, key points detection and facial attributes detection, recognizing the gender, age, race, emotion, face scores, fashion and other attributes of the characters in the images. It also supports the recognition of Chinese and foreign entertainment stars. The deep-learning based facial comparison technology achieves large-scale facial search and comparison, applicable to situations such as security.
Next, with the advantages of the company's existing technologies, Yi+ will continue to further explore the fields of smart cities, new retail, smart marketing, smart hardware and other fields so as to enable computers to understand the world, provide AI services and help people see the extraordinary and different.
Yi+ has received Series B financing from Alibaba and others and serves many Fortune 500 customers. Most of the team members come from top universities such as Columbia University, Imperial College London, Yale University, Princeton University, Purdue University, National University of Singapore, Nanyang Technological University, Tsinghua University, Peking University, and well-known companies such as Microsoft, IBM, Intel, Alibaba, Tencent, Baidu, Huawei and so on.
Related Links :