Search for a command to run...
Lightweight neural networks that employ depthwise convolution have a significant computational advantage over those that use standard convolution because they involve fewer parameters; however, they also require more time, even with graphics processing units (GPUs). We propose a Repetition-Reduction Network (RRNet) in which the number of depthwise channels is large enough to reduce computation time while simultaneously being small enough to reduce GPU latency. RRNet also reduces power consumption and memory usage, not only in the encoder but also in the residual connections to the decoder. We apply RRNet to the problem of resource-constrained depth estimation, where it proves to be significantly more efficient than other methods in terms of energy consumption, memory usage, and computation. It has two key modules: the Repetition-Reduction (RR) block, which is a set of repeated lightweight convolutions that can be used for feature extraction in the encoder, and the Condensed Decoding Connection (CDC), which can replace the skip connection, delivering features to the decoder while significantly reducing the channel depth of the decoder layers. Experimental results on the KITTI dataset show that RRNet consumes 3.84x less energy and 3.06x less memory than conventional schemes, and that it is 2.21x faster on a commercial mobile GPU without increasing the demand on hardware resources relative to the baseline network. Furthermore, RRNet outperforms state-of-the-art lightweight models such as MobileNets, PyDNet, DiCENet, DABNet, and EfficientNet.