Contents

Soft-GatedWarping-GAN for Pose-Guided Person Image Synthesis

https://gitee.com/shilongshen/image-bad/raw/master/20200707165658.png

target parsing:在特定的segmentation 区域进行纹理的渲染

warping block :align the image features

通过仿射变换TSP计算出transformation map用于warping condition image以减缓姿态不对齐问题


https://gitee.com/shilongshen/image-bad/raw/master/20200707164159.png

第一阶段生成目标姿态的part segmentation map

第二阶段首先训练geometric matcher 来估计condition segmentation和synthesized segmentation的transformation 参数。基于这个参数将condition image 的feature map进行warping,并渲染到target segmentation map

https://gitee.com/shilongshen/image-bad/raw/master/20200707164926.png

warping GAN的好处:1)如果存在大的姿态变换就会进行大的transformation,小的姿态变形就会进行小的transformation

2)在feature-level进行warping,能够合成更加真实的图片

3)warping block 能够通过attention layers自动选择有效的feature map 进行warping


Stage I: Pose-Guided Parsing

C:\Users\HIT\AppData\Roaming\Typora\typora-user-images\image-20200707183807443.png

为了在part-level上学习condition image到target pose之间的映射,引入了pose-guide parser.https://gitee.com/shilongshen/image-bad/raw/master/20200707184229.png


https://gitee.com/shilongshen/image-bad/raw/master/20200707184959.png

Geometric Matcher

https://gitee.com/shilongshen/image-bad/raw/master/20200707185450.png

将仿射变换和TPS相结合,通过孪生卷积神经网络来获得transformation map。

首先估计condition segmentation 和synthesized segmentation之间的仿射参数,基于仿射变换参数,估计经过仿射变换后的warping result和target parsing之间的TPS参数。

提取出来的transformation map用于warping条件图像的feature,减缓不对齐问题。

https://gitee.com/shilongshen/image-bad/raw/master/20200707190116.png

Soft-gatedWarping-Block

https://gitee.com/shilongshen/image-bad/raw/master/20200707205604.png

https://gitee.com/shilongshen/image-bad/raw/master/20200707205708.png


网络结构

https://gitee.com/shilongshen/image-bad/raw/master/20200707210653.png


消融实验

https://gitee.com/shilongshen/image-bad/raw/master/20200707211414.png