Deep Discrete Wavelet Transform Network for Photometric Stereo

Photometric stereo aims to estimate the per-pixel surface normal map of 3D objects via changing the illuminated light directions. Prevalent methods adopt deep neural networks to extract the shading cue features and reconstruct the surface normals. However, previous methods do not consider the frequency of the surface structure, i.e., the complexity of the shape. Simply applying a trained network to all kinds of objects often leads to inter-frequency conflicts and blur in surface normal estimation. This paper presents a discrete wavelet transform-based photometric stereo network (DWTPS-Net) to handle the input photometric stereo images in both the spatial and frequency domains. In DWTPS-Net, we extract shading features from images and also decompose the images using discrete wavelet transform (DWT), which can preserve spatial information naturally, to better extract high-frequency information. We design separate CNN-based feature-extraction modules for the input images and for the different frequency information of the input images via DWT. Ablation studies and experiments on a widely used benchmark dataset show that DWTPS-Net achieves superior performance in surface normal estimation, in terms of mean angular error metric.