一些看过的论文，防止遗忘

个人理解不一定对，欢迎讨论批评指正

Image Harmonization

1、DoveNet 【DoveNet: Deep Image Harmonization via Domain Verification】
paper：https://arxiv.org/pdf/1911.13239.pdf
公开了数据集Iharmony4。生成对抗网络，生成器：类似U-Net的encoder-decoder卷积网络；判别器：全局判别器+域验证判别器
2、iDIH 【Foreground-aware Semantic Representations for Image Harmonization】
paper：https://arxiv.org/pdf/2006.00809.pdf
encoder-decoder结构，预先提取语义信息辅助
3、S2AM 【Improving the Harmony of the Composite Image by Spatial-Separated Attention Module】
paper：https://arxiv.org/pdf/1907.06406.pdf
S2AM是个Attention模块
4、DIH 【Deep Image Harmonization】
paper：https://openaccess.thecvf.com/content_cvpr_2017/papers/Tsai_Deep_Image_Harmonization_CVPR_2017_paper.pdf
encoder-decoder结构，场景解析decoder分支辅助
5、Hao et al. 【Image Harmonization with Attention-based Deep Feature Modulation】
paper：https://www.bmvc2020-conference.com/assets/papers/0121.pdf
也设计了attention模块，the attention-based foreground-background feature map modulation (FMM)
6、RealismCNN 【Learning a Discriminative Model for the Perception of Realism in Composite Images】
paper：https://arxiv.org/pdf/1510.00477.pdf
用CNN评估图片真实性
7、SSH 【SSH: A Self-Supervised Framework for Image Harmonization】
paper：https://arxiv.org/abs/2108.06805
3D_LUTs + self-supervised
8、RainNet (#) 【SSH: A Self-Supervised Framework for Image Harmonization】
paper：https://arxiv.org/abs/2106.02853

Painterly Harmonization

1、E2STN (#) 【Element-Embedded Style Transfer Networks for Style Harmonization】
paper：https://bmvc2019.org/wp-content/uploads/papers/0425-paper.pdf
2、Zhang et al. (#) 【Deep Image Blending】
paper：https://openaccess.thecvf.com/content_WACV_2020/papers/Zhang_Deep_Image_Blending_WACV_2020_paper.pdf

Shadow Generation

1、LISA 【Instance Shadow Detection】
paper：https://openaccess.thecvf.com/content_CVPR_2020/papers/Wang_Instance_Shadow_Detection_CVPR_2020_paper.pdf
SOBA数据集：成对的物体和阴影标注。
2、ARShadowGAN 【ARShadowGAN: Shadow Generative Adversarial Network for Augmented Reality in Single Light Scenes】
paper：https://openaccess.thecvf.com/content_CVPR_2020/papers/Liu_ARShadowGAN_Shadow_Generative_Adversarial_Network_for_Augmented_Reality_in_Single_CVPR_2020_paper.pdf
数据集Shadow-AR，real-world image作为background，3D model为foreground来生成shadow free image，然后通过计算光源和相机位置来生成foreground的阴影（？）。
3、SSN 【SSN: Soft Shadow Network for Image Compositing】
paper：https://arxiv.org/pdf/2007.08211.pdf
如何生成物体的soft shadow（？）

Color Transfer

1、3D LUTs 【Learning Image-adaptive 3D Lookup Tables for High Performance Photo Enhancement in Real-time】
paper：https://arxiv.org/abs/2009.14468
学习一组3D LUT的权重系数，同时也对LUT进行更新
2、Spatial-aware 3D LUTs 【Real-time Image Enhancer via Learnable Spatial-aware 3D Lookup Tables】
paper：https://arxiv.org/abs/2108.08697
T*M个3D LUT，学习局部信息（pixel-wise category）
3、WB emulation frameworks 【What Else Can Fool Deep Learning?
Addressing Color Constancy Errors on Deep Neural Network Performance】
paper：https://openaccess.thecvf.com/content_ICCV_2019/papers/Afifi_What_Else_Can_Fool_Deep_Learning_Addressing_Color_Constancy_Errors_ICCV_2019_paper.pdf
transformation matrix

Style Transfer

1、AdaIN 【Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization】
paper：https://arxiv.org/abs/1703.06868

Photo Enhancement

1、HDRNet 【Deep Bilateral Learning for Real-Time Image Enhancement】
paper：https://groups.csail.mit.edu/graphics/hdrnet/data/hdrnet.pdf

Tricks

1、F2GAN 【F2GAN: Fusing-and-Filling GAN for Few-shot Image Generation】
paper：https://arxiv.org/abs/2008.01999v2
a mode seeking loss
1、BicycleGAN 【F2GAN: Fusing-and-Filling GAN for Few-shot Image Generation】
paper：https://arxiv.org/abs/1711.11586v1
‘This helps prevent a many-to-one mapping from the latent code to the output during training, also known as the problem of mode collapse, and produces more diverse results.’
a latent code vector z (e.g. a standard Gaussian distribution N(0, I))