NetEase Fuxi won the CVPR 2023 UG2+ and VizWiz competitions, and his paper was selected as TIP

Recently, the results of the CVPR 2023 competition were announced. NetEase Fuxi Lab achieved first place in the CVPR 2023 UG2 Haze Target Recognition Challenge and VizWiz Few-Sample Target Recognition Challenge. Their related papers have also been accepted by TIP, the top international journal. This shows that NetEase Fuxi's top technological innovation capabilities in the field of computer vision have been highly recognized internationally.

From February to June 2023, IEEE/CVF Computer Vision and Pattern Recognition Conference (CVPR), as the top conference in the field of international computer vision and pattern recognition, cooperates with authoritative academic institutions and well-known enterprises around the world , held a number of challenges. These challenges have attracted widespread participation from many AI research teams. Recently, CVPR has successively announced the award results and issued award certificates. As a top AI academic conference hosted by IEEE, CVPR has extremely high academic influence and social recognition.

In the CVPR 2023 UG2 Object Detection in Haze Challenge and the CVPR 2023 VizWiz Few-Shot Object Recognition Challenge, NetEase Fuxi teamed up with teacher Yu Jun from the University of Science and Technology of China and achieved first place. This cooperation mainly focuses on two aspects: target detection and few-sample target recognition in the field of computer vision. These technologies can be widely used in vision tasks in various fields. Especially in industrial applications, few-sample target detection is of great value and significance in scenarios where data acquisition and annotation are difficult. Through the success of this competition, we have demonstrated NetEase Fuxi’s research strength and innovation capabilities in the field of computer vision. We will continue to be committed to promoting the development of computer vision technology and providing more accurate and efficient solutions for practical applications.

The goal of UG2 is to advance the analysis of "difficult" images by applying image restoration and enhancement algorithms to improve analysis performance. Contestants are tasked with developing new algorithms to improve the analysis of images captured under problematic conditions. VizWiz's goal is to make more people aware of the technology needs and interests of people with visual impairments and to encourage artificial intelligence researchers to develop new algorithms to remove accessibility barriers. Competitions typically include tasks such as identifying objects in images, identifying text in images, and answering questions about images. The following is a brief overview of NetEase Fuxi’s award-winning paper:

Full-frequency Channel-selection Representations for Unsupervised Anomaly Detection

Omni-frequency Channel-selection Representations for Unsupervised Anomaly Detection

Keywords: Unsupervised image anomaly detection

Anomaly detection plays an important role in visual image understanding and is used to determine whether a given image deviates from a preset normal state. It is widely used in novelty detection, industrial image-based product quality monitoring, automatic defect repair, human health monitoring and video surveillance. Currently, there are three main types of mainstream unsupervised anomaly detection methods, including density-based methods, classification-based methods and reconstruction-based methods. These methods achieve anomaly detection by analyzing the statistical characteristics of images, learning normal samples, and reconstructing images, providing reliable tools and technical support for various applications.

Among them, the reconstruction-based method is rarely mentioned due to poor reconstruction ability and low performance, but it does not require a large amount of additional training samples for unsupervised training. More practical in industrial applications. To this end, this study focuses on improving the reconstruction-based method and proposes a new full-frequency channel selective reconstruction network (OCR-GAN), which is the first to handle the sensory anomaly detection task from the perspective of frequency. A large number of experiments have proved the effectiveness and superiority of this method compared to other methods. For example, without additional training data, new SOTA performance is achieved on the MVTec AD dataset, with an AUC of 98.3, significantly exceeding the reconstruction-based method baseline of 38.1 and the current SOTA method by 0.3.

The paper proposes an innovative solution to solve the UI anomaly problem in smart game compatibility testing. This solution uses artificial intelligence technology to automatically detect UI anomalies that occur when the game is running, and realizes the automation of game compatibility testing. By using image anomaly detection technology, we automatically detect a large number of generated game interface screenshots from the perspective of computer vision, obtain UI abnormal pictures from them, and assist game developers to quickly and accurately locate the cause of the problem, thus effectively saving game testing. The labor cost of experts.

This paper, in collaboration with the team of Professor Liu Yong of Zhejiang University, was selected for publication in the IEEE Transactions on Image Processing (TIP) journal. TIP is the top journal in the field of image processing research under IEEE. It is a journal in the SCI area of ​​the Chinese Academy of Sciences, and a Category A journal in the field of computer graphics and multimedia (CCF A) recommended by the China Computer Society. The journal's impact factor in 2022-2023 reaches 11.041.

The above is the detailed content of NetEase Fuxi won the CVPR 2023 UG2+ and VizWiz competitions, and his paper was selected as TIP. For more information, please follow other related articles on the PHP Chinese website!