VarGNet: Variable Group Convolutional Neural Network for Efficient Embedded Computing
This addresses efficiency challenges for embedded systems in computer vision, though it appears incremental as it modifies an existing convolution technique.
The paper tackled the problem of inefficient embedded computing in neural networks by proposing VarGNet, which fixes the number of channels in group convolutions instead of group numbers, resulting in easier hardware optimization and demonstrated practical value across multiple vision tasks.
In this paper, we propose a novel network design mechanism for efficient embedded computing. Inspired by the limited computing patterns, we propose to fix the number of channels in a group convolution, instead of the existing practice that fixing the total group numbers. Our solution based network, named Variable Group Convolutional Network (VarGNet), can be optimized easier on hardware side, due to the more unified computing schemes among the layers. Extensive experiments on various vision tasks, including classification, detection, pixel-wise parsing and face recognition, have demonstrated the practical value of our VarGNet.