A comprehensive exploration of depthwise separable convolutions for real time 3D hand pose estimation through RGB images

Abstract

Hand pose estimation is an important task in computer vision due to its various fields of application, but mainly for providing a natural interaction between humans and machines. There are significant challenges for solving this task, primarily due to the high degree of freedom that is present in the human hand, and the possibility of self-occlusion. We investigate the usage of depthwise separable convolutions, an optimized convolution operation, to speed-up the inference time for convolutional models trained for 3D hand pose estimation. We show that the execution time for this approach can be improved to be up to 34.28% faster, while maintaining the accuracy scores on the metrics proposed by the literature. Additionally, we performed an extensive exploration and analysis of the use of depthwise separable convolutions regarding common challenges in tracking such as blur and noise, aiming to understand better in which scenarios this type of convolution impacts on the tracker precision.