读书笔记PVANET: Deep but Lightweight Neural Networks for Real-time Object Detection

转自：http://www.cnblogs.com/xueyuxiaolang/p/5959442.html

传送门：

     论文：https://www.arxiv.org/pdf/1608.08021v3.pdf

     code：https://github.com/sanghoon/pva-faster-rcnn

中心：minimizing the computational cost 

目前通用的检测方法：（类似faster-rcnn）

the common pipeline of “CNN feature extraction + region proposal + RoI classification”

这篇论文在CNN feature extraction上做文章，设计原则是“less channels with more layers”，即更深更窄

adoption of some building blocks including concatenated ReLU, Inception, and HyperNet

trained with the help of batch normalization, residual connections

learning rate scheduling based on plateau detection