pifu code

[z-z-zhao]

您好，我看了您的ICON的仓库，其中关于PIFu的代码。我有了一些疑问，在2019年的PIFu论文中并没有使用法线的特征，为什么关于PIFu的您的代码中有法线的使用？ 另外您还使用了smpl的特征？（PIFu论文中并没有使用smpl）如下图：

恳请您帮助解决我的疑问！十分感谢！！！

[z-z-zhao]

2019年的PIFu(https://github.com/shunsukesaito/PIFu)

[z-z-zhao]

PIFu: Pixel-Aligned Implicit Function for High-Resolution Clothed Human Digitization

[YuliangXiu]

In my re-implemented PIFu, called PIFu* in ICON's paper:

• We do use normal maps to better align with other re-implemented methods, e.g., PaMIR, ICON. In this way, we can claim that the performance boost mainly comes from different methodologies (z-value vs. voxelized SMPL vs. local SMPL feature), instead of the input types.
• But PIFu* does not extract local features from SMPL (please go deeper into HGPIFuNet.py#L126), it only uses image+normal as input.

[z-z-zhao]

in your "But PIFu does not extract local features from SMPL (please go deeper into HGPIFuNet.py#L126), it only uses image+normal as input". is PIFu(your code) or PIFu (original paper) ?

As far as I know, PIFu* (original paper) did not use normals and it just use image,but PIFuHD used forward and backward normals + image. This is very important to me and I look forward to your reply！

[YuliangXiu]

Sorry for the confusion, I updated the last reply, PIFu* is from ICON's paper.

PIFu (original): use image as input PIFuHD: use image+normals(front+back, estimated from image) as input PIFu* (ICON's re-implementation): use image+normals(front+back, estimated from image & conditioned on SMPL-X) as input