From Shadow Generation to Shadow Removal （CVPR2021）阅读笔记

前人的对于阴影的数据集的采用：
For most of them a set of paired shadow images and their corresponding shadow-free version are used to train the network in a fully supervised manner
其中大部分是使用一组成对的阴影图像及其对应的无阴影版本，以完全监督的方式训练网络

they leverage the fact that a shadow image usually contains both shadow and non-shadow regions. This way, a set of shadow and shadow-free patches can be cropped to construct unpaired data for network training ECCV2020
他们利用了阴影图像通常包含阴影和非阴影区域这一事实。通过这种方法，可以裁剪一组阴影和无阴影的补丁来构建不配对的数据，用于网络训练

本文的方法：
Given an input shadow image, the shadow-generation sub-net generates pseudo shadows for each shadow-free region and such pseudo shadows are then paired with the corresponding original shadow-free region to form the training data
给定一幅输入阴影图像，阴影生成子网络对每个无影区域生成伪阴影，然后将伪阴影与对应的原始无阴影区域配对，形成训练网络

we propose a new G2R-ShadowNet that consists of three sub-networks for shadow generation, shadow removal and refinement, respectively
提出三个子网模块：阴影产生，阴影移除，阴影细化

具体思想流程

损失函数
By combining all the loss functions proposed for the above three sub-nets, the total loss L for training the shadow generator G, the shadow removal sub-net I and the refinement sub-net R is defined as:
将上述三个子网提出的所有损失函数结合起来，定义训练阴影发生器G、阴影去除子网I和细化子网R的总损失L为:

实验部分：
The ISTD dataset is proposed for both shadow detection and shadow removal and the data are collected under various illumination conditions with different shadow shapes
用于阴影检测和阴影去除的ISTD数据集，该数据集在不同光照条件和不同阴影形状下采集

Evaluation metrics
For all experiments, we use the Root-Mean-Square Error (RMSE), Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM) as the evaluation metrics.
在所有实验中，我们使用均方根误差(RMSE)、峰值信噪比(PSNR)和结构相似度(SSIM)作为评价指标。

消融实验
1 One is obtained by removing R and the loss functions that are related to R. The other one is obtained without using the shadow mask M as the input which means the region that is going to be refined is not known
一种是通过去掉R和与R相关的损失函数得到的，另一种是在不使用阴影掩模M作为输入的情况下得到的，这意味着需要细化的区域是未知的

2 we try to detach the result of each sub-net individually and train each variants one-by-one to see how one impacts the others . For instance, when the refinement result is detached, the back-propagated signal from the refinement sub-net R is not passed to G and I.
我们试图单独分离每个子网络的结果，并逐个训练每个变体，以了解一个变体如何影响其他变体。例如，当分离细化结果时，从细化子网R反向传播的信号不传递给G和I。

3 We also conduct another ablation study to justify the effectiveness of each loss function by training our model without a specific loss term for each time
我们还进行了另一项消融研究，通过训练我们的模型，每次都不使用特定的消融项来证明每个消融功能的有效性

对比效果图：

与他人的方法对比：
compare our proposed weakly supervised method with several state-of-the-art methods on ISTD

shows the qualitative results of our method and the other state-of-the-art methods on four challenging samples drawn from the testing set of ISTD