GoodFuture AI Institute Wins Four International Championships at CVPR 2021 Across Multiple Vision Challenges

After months of intense competition, the results of several CVPR 2021 challenges (Computer Vision and Pattern Recognition) were announced. GoodFuture AI Institute, leveraging long‑term investment and innovative technology in the image domain, won four international championships in the Person In Context, UG²+, ETH‑XGaze, and ActivityNet challenges, demonstrating that Chinese educational‑technology AI research has reached world‑leading levels.

Following its CVPR 2020 EmotioNet facial expression recognition championship, GoodFuture again achieved outstanding results at this top‑tier conference. The cutting‑edge technologies presented include low‑light face detection, human‑object interaction detection, gaze estimation, and active speaker detection.

Person In Context Challenge focuses on human‑object relationship detection, requiring output of triplets containing human boxes, object boxes, and their relationships. GoodFuture proposed a graph‑network method that incorporates human keypoint information for relationship reasoning and integrates a Transformer module for feature enhancement, achieving an absolute mAP advantage of 95.5 and applying the technology to its teaching‑quality assessment system.

UG²+ Challenge addresses the performance drop of visual perception algorithms under unconstrained, dynamically degraded conditions such as harsh weather and lighting. GoodFuture combined traditional and deep‑learning “de‑darkening” pipelines, using image‑processing and GAN techniques to simulate dark‑to‑bright transformations with added noise, ultimately surpassing the runner‑up by more than 3% mAP to win the competition.

ETH‑XGaze Challenge requires precise estimation of gaze direction from images, a key factor in human‑computer interaction, affective computing, and medical diagnosis. GoodFuture and the Institute of Computing Technology, Chinese Academy of Sciences, fused eye‑local features with global facial features via an attention mechanism, significantly improving regression accuracy. The technology has been deployed in the online teaching system of TAL Education Group.

ActivityNet Challenge (Active Speaker Detection) aims to determine whether visible persons in a video are speaking. Without any pre‑training, GoodFuture achieved a 93.44 mAP score, outperforming teams from Technical University of Munich, Microsoft, and National University of Singapore.

GoodFuture AI Institute, a pioneer in AI research for education in China, has accumulated nearly five years of breakthroughs in image and video understanding, machine learning, natural language processing, and speech synthesis/evaluation. It offers over 100 AI capabilities across vision, speech, NLP, and data mining, has published more than 60 high‑quality papers, and holds over 100 patents.

The institute is a founding partner of the National New‑Generation AI Open Innovation Platform for Smart Education and has recently partnered with Zhiyuan Research Institute to accelerate intelligent classroom, teacher‑assistant, smart companion, education‑hardware, and adaptive learning solutions.

Scan the QR code below to add the official WeChat account “GoodFuture Technology” for real‑time interaction, or add the WeChat ID TAL-111111 directly.

Also check out related articles: Service Mesh Practices in the Online School Platform , Deep Dive into Container Networking , and Upgrading the Web Live Classroom Architecture .