Ping An Insurance (Group) Company of China, Ltd. announced that the computer vision object detection model developed by Ping An Technology (Shenzhen) has set a new record in the PASCAL VOC Challenge, one of the most authoritative competitions in the world to assess the design and innovation capabilities of AI algorithms.
In the PASCAL VOC comp3 object detection challenge, Ping An Technology earned a mean average precision (mAP) of 86.5% and ranked first in 18 of 21 indicators, to take first place overall in a field of 59 competitors, including numerous notable artificial intelligence (AI) enterprises and AI laboratories in universities around the world. Object detection is the functionality to determine where there are the instances of a particular object class – such as a bird, a chair or a person – in an image, if any.
The Pattern Analysis, Statistical Modelling and Computational Learning Visual Object Classes (PASCAL VOC) Challenges have attracted the participation of numerous tech giants and top research institutions including Intel, Alibaba, Tencent and Microsoft Research since they were launched in 2006.
Object detection is one of the three basic tasks in the field of computer vision. The VOC2012 comp3 challenge has been one of the fiercest PASCAL VOC competitions, due to the diversity of target types and scenarios and the relatively small data set. These make identification difficult but most accurately evaluates the detection performance of different algorithms.
Ping An Technology’s NAS-YoLo model uses the company’s knowledge accumulated in automated machine learning and deep learning. It introduced AutoAugment and Neural Architecture Search (NAS) to the YoLo (You Only Look Once) model to significantly increase the precision of object detection. The model selects the most compatible data augmentation strategy for different data sets through AutoAugment, which resolves issues such as small data sets and difficult-to-enhance performance. By incorporating NAS, based on a super network, the solution speeds up the searching process with a divide-and-conquer algorithm, which reduces the time and difficulties of the search. To increase accuracy, the team also incorporated an automatic parameter tuning method based on Sequential Model Based Optimization (SMBO), to generate the best hyperparameter combination for NAS-YoLo, thus achieving a major breakthrough in the field of object detection.
The NAS-YoLo model is trained on an Occam automated machine learning platform developed by Ping An. The Occam platform is an AI algorithm platform that integrates six cutting-edge machine learning technologies: automatic pipeline, auto augmentation, distributed training acceleration, automatic model compression, auto-tuning, neural architecture search. With its comprehensive library of ground breaking algorithms for images, speech and text, the Occam platform makes multiple models, including NAS-YoLo and more than 100 kinds of other AI resources available to the public. The Occam platform is committed to helping industrial development in various fields with more advanced and scientific technologies.
The Occam platform has served dozens of internal research teams of Ping An in areas such as speech, voiceprint and facial recognition, Optical Character Recognition (OCR), Natural Language Processing (NLP) and medical imaging. It has accelerated the development of AI technologies in Ping An Property & Casualty, Ping An Life, Ping An Bank, Puhui Business and other Ping An Group companies.