Photometric stereo using CNN-based feature-merging network

We propose a photometric stereo method using Convolutional Neural Network (CNN) based method, which is effective for deriving surface normal data from non-lambertian objects. Our method extracts feature maps from a set of images of object using shared feature extraction network, and merge the extracted feature maps using two pooling method: max-pooling and average-pooling. The merged feature maps are concatenated and passed to final CNN layers to derive the surface normal map. We tested our network on the most widely-used benchmark dataset and confirmed that our method performs better than existing deep learning based photometric stereo method.