A Joint Training Framework Based on Adversarial Perturbation for Video Semantic Tags Classification

Download PDF
Open Webpage

Zhenxu Ye, Yaoyuan Zhang, Shengming Cao

Published: 2021, Last Modified: 05 Nov 2025CCKS (Evaluation Track) 2021EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: In this paper, we present a solution of “Knowledge Enhanced Video Semantic Understanding” of 2021 CCKS Track 14th task [1]. We separate video semantic understanding framework into two related tasks, namely the multi-classes video cate classification (VCC) task and the multi-label video tag classification (VTC) task. Meanwhile we propose a joint training framework for VCC task and VTC task based on adversarial perturbations strategy. In the final leaderboard, we achieved 3rd place in the competition. The source code has been at Github (https://github.com/stone-yzx/2021-CCKS-Trace14-3rd-semantic-tag-classification).