论文阅读笔记:“MGFN:Magnitude-Contrastive Glance-and-Focus Network for Weakly-Supervised Video Anomaly Detection”

论文阅读笔记:“MGFN:Magnitude-Contrastive Glance-and-Focus Network for Weakly-Supervised Video Anomaly Detection”
网络框架

核心亮点
Feature Amplification Mechanism (FAM)

对于第i个视频的第t个片段,FAM会计算一个特征范数,他是在特征维度进行归一化
之后使用1d卷积调制范数信息来进行增强信息导出:

Magnitude Contrastive Loss
之前经典方案是二元交叉熵损失:
论文提出了振幅对比损失:

p,q是正常clip的索引,u,v是异常clip的索引,Ma是异常视频top-k高的特征振幅。如果p,u是一对正常,异常视频clip

总体loss计算公式:

- Post title:论文阅读笔记:“MGFN:Magnitude-Contrastive Glance-and-Focus Network for Weakly-Supervised Video Anomaly Detection”
- Post author:sixwalter
- Create time:2023-08-05 11:14:26
- Post link:https://coelien.github.io/2023/08/05/paper-reading/paper_reading_065/
- Copyright Notice:All articles in this blog are licensed under BY-NC-SA unless stating additionally.
Comments