Paper reading：高分辨率图像分割 Collaborative Global-Local Networks for Memory-Efficient Segmentation CVPR19

现有的大尺度图像分割方法大多是下采样或者切分为图像块进行处理，但是会导致局部细节和全局纹理信息的缺失。因此本文提出协作式全局-局部网络(Collaborative global-local network)，以一个高效的方式保留全局和局部信息。协作式全局-局部网络有两个分支：全局分支和局部分支，分别处理降采样后的全局图像和剪切后的局部图像块作为输入信息。为了完成分割任务，网络会融合两个分支的特征图，以更好地获取高分辨率的细节结构和上下文依赖关系。此外，为了解决类别不均衡问题（即前景和背景区域不均衡），提出一个从粗到细的网络变体，具有内存-速度的双重优势。

网络架构：

从粗到细的网络框架：

实验细节：For simplicity, both the downsampled global image and the cropped local patches share the same size, 500*500 pixels. Neighboring patches have 50-pixel overlap to avoid boundary vanishing for all the convolutional layers.

 实验数据库：

（1）DeepGlobe dataset. This dataset contains 803 ultra-high resolution images (24482448 pixels). We randomly split images into training, validation and testing sets with 455, 207, and 142 images respectively. The dense annotation contains 7 classes of landscape regions, where one class out of seven called “unknown” region is not considered in the challenge. 

 （2）ISIC Lesion Boundary Segmentation Challenge dataset contains 2594 ultra-high resolution images. We randomly split images into training, validation and testing sets with 2074, 260, and 260 images respectively.

 （3）Inria Aerial Challenge dataset contains 180 ultrahigh resolution images, each with 50005000 pixels. We randomly split images into training, validation and testing
sets with 126, 27, and 27 images respectively.

 本文的两个贡献点/结论：

（1）We proposed a memory-efficient segmentation model GLNet specifically for the ultra-high resolution images. It leverages both the global context and local fine structure effectively to enhance the segmentation in the scenario of ultra-high resolution without sacrificing the GPU memory usage.

（2）We also proved that the class imbalance problem can be solved by our coarse-to-fine segmentation approach.