Real-Time High-Performance Semantic Image Segmentation of Urban Street Scenes解读

论文地址:https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9042876 年份:2020

关键词:轻量级、精度和速度的平衡

一、工作

1.提出了一个高性能实时分割网络,用于分割城市街道景色。实现了速度精度更好的平衡。上图:

Real-Time High-Performance Semantic Image Segmentation of Urban Street Scenes解读
和主流的一些网路的对比

 

2.用到的模块:

(1)distinctive atrous spatial pyramid pooling/ASPP,作用:提取不同size的pooling层的信息,让encoder的ferture更加丰富、易于分辨,用作检测多尺度的目标。

(2)spatial detail-preserving network/SPN,作用:带有浅卷积层,用来产生高分辨率的feature map,保护空间信息的细节。

(3)feature fusion network/FFN,作用:有效的结合来自DSPP和SPN的深层和浅层的feature。

(4)baseline:lightweight baseline network with atrous convolution and attention/LBN-AA

3.表现

cityscapes:73.6%miou(test),51.0fps。camid:68.0%miou(test),39.3fps

Real-Time High-Performance Semantic Image Segmentation of Urban Street Scenes解读
网络结构
Real-Time High-Performance Semantic Image Segmentation of Urban Street Scenes解读
网络的表现cityscapes