论文阅读笔记：“ActionVLAD：Learning spatio-temporal aggregation for action classification”

ActionVLAD: Learning spatio-temporal aggregation for action classification

代码：http://rohitgirdhar.github.io/ActionVLAD

论文思想

时空信息聚合

将特征空间划分为K个区域，该区域可以表示为“action words”，也可以称其为锚点(achor points ck)

公式：

上面的公式对特征与锚点（typical actions）之前的差异在整个视频维度进行了求和。

why use VLAD to pool?

HOW to combine RGB and FLOW streams?

Ideas?

• can I add a attention to VLAD?
• where to add attention?
• can I add non-local to VLAD?

Todos

• run the VLAD code
• look at the core codes where the VLAD is implemented
• try out the ideas that may work well