Frequency selective regularization methods for video model

Abstract

Recently, various regularization techniques for deep neural network based video models have been studied. From the perspective of supervised learning, it is known that deep neural network models for video action recognition are easily prone to overfitting to training data due to a large number of parameters, and regularization techniques can be one solution to alleviate the problem. From the perspective of unsupervised learning (or self-supervised learning), a group of data augmentation methods, which is a type of regularization method, is used in various ways as an essential element of the contrastive learning method that is being actively studied. In this study, we propose frequency-selective regularization techniques for supervised/unsupervised learning of video models. First, in order to solve the overfitting problem of the video action recognition model, we propose a regularization technique in which small random changes are made to low-frequency signals in the feature stage. Second, we propose a data augmentation method in which the video model arbitrarily filters the spatiotemporal low-frequency signal from the input video signal to learn a better representation through contrastive learning. Through these frequency-selective regularization techniques, it can be confirmed that the video model improves the action recognition performance in the target task without using additional training data.