모바일 환경에서의 실시간 손동작 인식을 위한 경량화된 3D-CNN

Abstract

Despite its high accuracy, 3D-CNN, which is neural network for action recognition, have not been able to be utilized in mobile processor due to massive computation and memory footprint of 3D convolution. To facilitate real-time operation of 3D-CNN for hand gesture recognition, this paper proposes two key features: 1) Inter-frame Differential Aware(IDA) input method to maximally utilize the activation sparsity; 2) IDA Network(IDANet) optimized for IDA input method for reducing network parameter. IDANet reduced 96% of parameters and achieved 83.2 activation sparsity throughout the network and 79.97% accuracy on NvGesture data set. As a result, IDANet suggest possibility of real-time operation of 3D-CNN for hand gesture recognition in a mobile platform.