Header Ads

OPPO Bags 11 Awards at the Computer Vision and Pattern Recognition Conference 2021


Leading global smartphone brand OPPO recently took part in the premier annual computer vision event Computer Vision and Pattern Recognition Conference (CVPR) 2021. During the conference, OPPO's achievements in AI were recognized with its placing in six major Challenges in eleven different contests in total demonstrating the company's industry-leading technological strengths and innovative breakthroughs in AI. The team participating in the CVPR 2021 competition on behalf of OPPO came from the Intelligent Perception and Interaction Department of the OPPO Research Institute, whose entries in the six major Challenges of Multi-Agent Behavior, SoccerNet, LOVEU, AVA-Kinetics, MMact, and People in Context ended with excellent results. Through the optimization and training of AI algorithms, the team's work continues to strengthen OPPO's AI capabilities and the ability of its AI technology to better serve people.

"We are very pleased to have achieved such remarkable results again in this year's CVPR Challenges, following our inaugural participation in CVPR 2020. Last year, we won first place in the Perceptual Extreme Super-Resolution Challenge by demonstrating technology that can sharpen the appearance of blurry images and in Visual Localization for Handheld Devices Challenge that makes fusion positioning more precise. The Challenges won by OPPO this year, such as Multi-Agent Behavior, AVA-Kinetics, and 3D Face Reconstruction from Multiple 2D Images, cover more complex and advanced areas of computer vision, including behavior detection, localization of human actions in space and time, and facial detection," said Eric Guo, Chief Scientist, Intelligent Perception of OPPO.  "These technologies can be used in a whole range of scenarios such as manufacturing, home, office, photography, health, and mobility. At OPPO, we are committed to making AI better serve people, providing users with more intelligent and convenient experiences.” he added.

Multi-Agent Behavior Challenge
Multi-Agent Behavior Challenge

Among its eleven honors, OPPO received three awards in the Multi-Agent Behavior Challenge, which assesses an AI model's ability to understand, define, and predict complex interactions between intelligent agents such as animals and human beings. OPPO ultimately won the first-place prize in the Learning New Behavior category, second-place in Classical Classification, and third-place in Annotation Style Transfer, distinguishing itself from over 240 other participants thanks to its leading AI capabilities. This same technology is currently playing an essential role in OPPO's factory, where the algorithms assist workers in reducing operational mistakes, particularly in key production steps, to ensure their own safety as well as the quality of products coming off the production line.

From computational intelligence to human-centric intelligence, OPPO improves AI's ability to understand people
Through its mission of "Technology for Mankind, Kindness for the World," OPPO is building capabilities in human-centric AI. In the 3D Face Reconstruction from Multiple 2D Images Challenge, OPPO's self-developed AI algorithm was able to reconstruct 3D facial shapes with an error of around 1mm, leading it to take second place in the main index score ranking. OPPO's technology overcomes problems associated with unclear facial features, exaggerated expressions, and even damaged image data caused by real-life movements, especially in dynamic videos, to produce more accurate 3D facial models.

OPPO's self-developed facial detection algorithm is able to identify 635 key features points at a rate of 30 times per second. This same algorithm architecture is used to power the AI makeup video feature found in the upcoming OPPO Reno6 smartphone, which allows users to easily create dynamic and natural beauty effects in their videos. This technology will promote the evolution of portrait video technology, with 3D feature recognition making makeup effects and filters appear more lifelike and personalized.

3D Face Reconstruction from Multiple 2D Images
3D Face Reconstruction from Multiple 2D Images

AI that understands space and time
OPPO's AI capabilities have already developed to the stage where they are able to recognize human actions in space and time. In the SoccerNet Challenge, OPPO took second place in both the Action Spotting and Replay Grounding tasks. The purpose of the challenge was to evaluate the ability of the algorithms to identify more than a dozen key actions in a video of a soccer game, including offside and red card violations, which are usually difficult for humans to recognize due to the complexity of the rules and the subtle interpretations of them. To be effective, the AI algorithm also needs to account for other variables such as different camera angles, as well as accurately retrieve the timestamp of the action shown in a given replay shot within the original game. The future applications of this technology are wide-reaching and will help improve the experience for sports lovers through features such as automatically-generated match highlights. In a similar fashion, the technology can also be used to automatically create highlights of a user's life – for example weekly highlight clips – by analyzing the videos on their smartphone.

SoccerNet Challenge
SoccerNet Challenge

In the MMact Challenge, OPPO took second place in both the Cross-Modal Action Recognition and Cross-Model Action Temporal Localization tasks. OPPO's powerful AI algorithm can accurately recognize more than ten types of actions in a video, such as talking, crouching, and walking, using only visual data. This technology is expected to be widely adopted in smart homes in the future, with benefits including the ability to better take care of the children, pets, the elderly, or other vulnerable groups at home. For example, the AI can alert parents in another room as soon as a baby or child exhibits actions that could be potentially dangerous.

OPPO also won third place in the AVA-Kinetics Challenge, which makes use of the industry's first dataset to include both space and time information. The Challenge's Positioning competition has long been one of the most popular competitions in the field of artificial intelligence, with competitors including those from top international technology companies and universities. The AVA-Kinetics algorithm can not only accurately identify the various behavior of people in the video, but also note their time and position.

OPPO continues to explore the frontiers of AI technology
At this year's CVPR, OPPO also made a new milestone in the more cutting-edge academic challenges, including taking two third-place in LOVEU (Long-form Video Understanding) challenge. The LOVEU challenge requires AI technology to understand the content of a video and segment it into chunks without being given pre-defined categories. Given the huge possible variety of content, the challenge poses a significant test for the ability of AI algorithms to be applied to more generalized situations: The AI needs to think like a human being, understand colors, objects, human actions, and even light in the video, and make judgments on how these change over time. In the future, this technology has the potential to be widely used as the foundation for further AI tasks in video processing, such as facial detection and behavior recognition.

No comments:

Powered by Blogger.