SOD算法：BASNet 2019 CVPR

->BASNet
Xuebin Qin, Zichen Zhang, Chenyang Huang, Chao Gao, Masood Dehghan and Martin Jagersand University of Alberta, Canada
BASNet: Boundary-Aware Salient Object Detection

评测对比：


1 结构
预测+精炼
Encode-decode + novel residual refinement module

2 混合loss（三合一）
BCE: Binary Cross Entropy
SSIM: Structure SIMilarity
IoU: Intersection-over-Union

3 Encoder-Decoder
Encoder-Decoder network transfers the
input image to a probability map,
while the refinement module refines the
predicted map
by learning the residuals between the coarse saliency map and ground truth

->Introduction 要注意2点
i)
In contrast to [50, 22, 6], which use refinement modules iteratively on saliency predictions or intermediate feature maps at mul-tiple scales, our module is used only once on the original scale for saliency prediction.
不同于[50, 22, 6]，在多尺度对中间层或预测图迭代优化，
本文只对原始尺寸迭代一次。[6] is R3Net
ii)
BCE: Binary Cross Entropy
SSIM: Structure SIMilarity
IoU: Intersection-over-Union

分别学习了pixel级别，patch级别，map级别信息。
which are expected to learn from ground truth information in pixel-, patch- and map- level, respectively.

并非NLDF C2S显式的边界损失，隐式的注入目标的准确边界预测，
考虑到这可能有助于减少从交叉传播的边界和图像上的其他区域学习到的信息的虚假错误。
Rather than using explicit boundary losses (NLDF+ [41], C2S [36]), we implicitly inject the goal of accurate boundary prediction in the hybrid loss,
contemplating that it may help reduce spurious error from cross propagating the information learned on the boundary and the other regions on the image.

->contribution
i)编码解码、残差优化模块
En-de and residual refinement model
ii)三个损失合一

->related work
1 Traditional Methods:
Handcrafted feature [69,80,60,71]
Borji 的综述 comprehensive survey [3]
2 Patch-wise Deep Methods:
在多尺度，对图像像素分类或超像素分类，
因为全链接层的空间信息损失而产生粗糙输出
classifying image pixels or super pix- els into salient or non-salient classes based on the local image patches extracted from single or multiple scales [33 40 61 79 35]
3 FCN-based Method:
UCF、Amulet、NLDF、DSS、HED、RAS、LFR、BMPM
4 Deep Recurrent and attention Methods:
PAGRN、RADF、RFCN、PiCANetR
5 Coarse to Fine Deep Methods:
SRM、R3Net、DGRL

->prediction module:
is a U-Net-like densely supervised Encoder-Decoder network [57]
->The multi-scale Residual Refinement Module (RRM):
 refines the resulting saliency map of the prediction module by learning the residuals between the saliency map and the ground truth


the last layer of each decoder stage is supervised by the ground truth inspired by HED
每阶段末尾都有监督

encoder
一盒输入卷基层+6stage 基本的res-blocks
前4stage采用rennet-34,不同的是输入层是3x3的64卷基层而不是7x7步长2的
64 convolution filters with size of 3×3 and stride of 1 rather than size of 7×7 and stride of 2.
卷积后没有池化
there is no pooling operation after the input layer.