Zhang Yuanhang ; Liang Susan ; Yang Shuang ; Shan Shiguang

Abstract

This report presents a brief description of our winning solution to the AVAActive Speaker Detection (ASD) task at ActivityNet Challenge 2022. Ourunderlying model UniCon+ continues to build on our previous work, the UnifiedContext Network (UniCon) and Extended UniCon which are designed for robustscene-level ASD. We augment the architecture with a simple GRU-based modulethat allows information of recurring identities to flow across scenes throughread and update operations. We report a best result of 94.47% mAP on theAVA-ActiveSpeaker test set, which continues to rank first on this year'schallenge leaderboard and significantly pushes the state-of-the-art.