This project uses ThunderNet as detection framework, DlaNet as backbone network, ShuffleNetV2 block as lightweight module, papers and codes i use for reference are shown below.
ShuffleNetV2 paper
This project gets better performance when facing small objects because DlaNet fuses features better and gets larger feature map. At the same time, its computation complexity and run time decrease greatly as lightweight framework and convolution module.
Ubuntu16.04 GTX 1080 Ti Pytorch 1.0 cuda 9.0
This project is mainly based on ThunderNet-ouyanghuiyu github, so you can view the webpage above to use this project as a whole. By the way, A related blog will show you how to use this project to train and test in detail. It will be finished soon...
blog here introduce the project in detail and show you the result pictures.