Dimensionality Deduction for Action Proposals: To Extract or to Select?

Action detection is an important item in machine vision. Currently some works based on deep learning framework have achieved impressive accuracy, however they still suffer from the problem of low speed. To solve this, researchers have introduced many feasible methods and temporal action proposals method is the most effective one. Fed with features extracted from videos, these methods will propose time clips which may contain actions to reduce computational workload. It is a common way to use 3D convolutions (C3D) to extract spatio-temporal features from videos, nonetheless the dimension of these features is generally high, resulting sparse distribution in each dimension. Thus, it is necessary to apply dimension reduction method in the process of temporal proposals. In this research, we experimentally find that in action detection proposal task, reducing the dimension of features is important. Because it cannot only accelerate the process of subsequent temporal proposals but also makes its performance better. Experimental results on the THUMOS 2014 dataset demonstrate that the method of feature extraction reduction is more suitable for temporal action proposals than feature selection method.