Dynamic Convolution: Attention over Convolution Kernels

[jinglescode]

Paper

Link: https://arxiv.org/pdf/1912.03458.pdf
Year: 2020

Summary

• increases model complexity without increasing the network depth or width
• single convolution kernel per layer, dynamic convolution aggregates multiple parallel convolution kernels dynamically based upon their attentions, which are input dependent
• can be easily integrated into existing CNN architectures

Video: https://www.youtube.com/watch?v=FNkY7I2R_zM

Contributions and Distinctions from Previous Works

• two most popular strategies to boost the performance are making neural networks “deeper” or “wider”. However, they both incur heavy computation cost, thus are not friendly to efficient neural networks.
• dynamic convolution, which does not increase either the depth or the width of the network, but increase the model capability by aggregating multiple convolution kernels via attention

Methods

Results

• comparing to classic MobileNetV2, MobileNetV3 and ResNet, with dynamic convolution, it significantly improves the representation capability with negligible extra computation cost
• simply replacing each convolution kernel in MobileNet (V2 and V3) with dynamic convolution, we achieve solid improvement for both image classification and human pose estimation

Comments