动作识别-Regularization on Spatio-Temporally Smoothed Feature for Action Recognition-CVPR2020

Abstract

3D 卷积核，由于参数量多，容易overfitting；提出了一种正则化方法；the key idea of RMS is to randomly vary the magnitude of low-frequency components of the feature to regularize the model.

Introduction

解决overfitting问题有Perturbation base regularization methods on the input space and on the feature space; describe direction to perturb; 有关频域信号的提取这样定义

图像处理之高斯滤波的几种实现方式
The validation accuracy drops more quickly when high-freq. components are perturbed.
To separate the low-frequency feature, we use a 3D mean filter (3D average pooling operation in most deep learning frameworks) or a Gaussian filter which are the simplest low-pass filters (LPF) in image processing《Feature Extraction &Image Processing for Computer Vision第三版；2012》;
The authors’ intuition is that high-frequency component may possess more essential information for classification while low-frequency component may contain peripheral information so that small perturbation on it adds diversity of the sample without changing the type of action or object.

Method

实现部分要注意细节

Conclusion

提出了一种正则化方法RMS for 3D ResNet

key points: 关键理解如何定义low-frequency和high-frequency component；对滤波器的工程实现；文章的实验布局分析也可以学习一下；高频信号扰动，对分类性能影响大；对低频信号扰动通过乘上a single random scalar \alpha（从一个高斯分布中采样得到的\alpha）