The supplement training dataset for self-calibrating photometric stereo

The PS_Sculpture training dataset introduced by the PS-FCN contained various non-Lambertian reflectances, cast shadows, interreflections and effective noise information. However, for dark materials such as black-phenolic and steel, significant data loss happened due to 8-bit quantification. To lessen this data loss, we designed a new supplementary training dataset rendered by 10 blobby objects and 10 other objects freely downloaded from the Internet and the real BRDF data comes from the MERL dataset. Besides, this supplementary dataset includes the SVBRDF and the specific image illuminated by the collocated light. Better deep learning based photometric stereo model could be trained by the combination of the PS_Sculpture dataset and the new supplementary dataset.

PS-FCN引入的PS_Sculpture训练数据集包含多种非朗伯反射（non-Lambertian reflectances）、投射阴影、相互反射及有效噪声信息。然而，对于黑色酚醛（black-phenolic）和钢材等深色材质，8位量化（8-bit quantification）导致了显著的数据丢失。为减轻此类数据丢失，我们设计了一个新的补充训练数据集——该数据集由10个blob状物体和10个从互联网免费下载的其他物体渲染生成，且真实双向反射分布函数（BRDF）数据来源于MERL数据集。此外，该补充数据集包含空间变化双向反射分布函数（SVBRDF）及由共置光源照射的特定图像。将PS_Sculpture数据集与新补充数据集结合，可训练出性能更优的基于深度学习的光度立体模型（photometric stereo model）。