Accuracy-efficiency Balanced Sampling Rate Found by Discrete Fourier Transform
Jiongliang Lin, Yahui Peng, Yufei Yang, Xianhua Dai · 2022 7th International Conference on Signal and Image Processing (ICSIP) · 2022
Though Vision-based human action recognition has received extensive research, there are few works to explore the sampling rate of videos, which is essential in practical application. Too high sampling rates result in waste of computing resources and reduced efficiency, while too low sampling rates cause a rapid drop in accuracy. In this paper, we propose a method based on Discrete Fourier Transform (DFT) to find out a sampling rate which can balance efficiency and accuracy. This method captures the variation of video features in the frequency domain, which helps to choose a sampling rate based on how much information is required. We propose a 3D convolutional neural network for arbitrary input length (AC3D) and conduct extensive experiments on two widely used human action recognition datasets: UCFIOI and HMDB51. Experimental results show that DFT-based method can find a sampling rate with a good balance of efficiency and accuracy.