the first attempt to infer light occlusion in the SH formulation directly

infer not only an albedo map, illumination but also a light transport map that encodes occlusion as nine SH coefficients per pixel.

architecture

image-20220819093118355

为什么light transport map $\in \R^{NX9}$,N是pixel数量

15 L1 loss functions??

light transport map是什么

和SfSNet很接近

encoder有六层卷积，输出为{64, 128, 256, 512, 512, 512}，步幅stride为2

Each decoder has a residual block (consisting of two convolutional layers with 512 channels) and six deconvolutional layers (output channels are { 512, 512, 256, 128, 64, 9 or 3 } and the stride is also two).

Encoder、Albedo decoder的ResNet block、Transport decoder的ResNet block的结果连接起来传入四层卷积中，产生27维向量。

每个卷积层和反卷积层（除了第一个和最后一个）后都要做batch normalization和(leaky) ReLU

前三个反卷积层dropout概率为0.5

optimizer用Adam

lr为0.0002

batch size 1??

GTX 1080上跑一个epoch需要三个小时。。。，共60个epoch

数据集
Synthetic human image dataset

一部分来源于公开的BUFF，一部分买来。

从BUFF中挑出来74个代表性的模型

买了271个模型

276用于训练，69进行测试

aligned 3D models(脸朝前，垂直方向大小一致，padding一致（上下padding都为图片的5%）)。

只用了战力的人物

图片分辨率1024*1024

Illumination dataset

Laval Indoor HDR dataset

image-20220819100442181