Demultiplexing Colored Images for Multispectral Photometric Stereo via Deep Neural Networks

Recovering fine-scale surface shapes is a challenging task in computer vision. Multispectral photometric stereo is one of the popular methods as it can handle non-rigid/moving objects and produces per-pixel dense results. However, the colored images captured by practical multispectral photometric stereo setups are aliased in RGB channels. Existing solutions require prior information to calibrate few points and estimates whole surface normal by the calibration, while prior information is not always available and accurate. Differing from previous solutions which require calibration or other prior information, we first formulate the problem in a learning framework, which directly seeks the per-pixel mapping of the aliased and spectrum-multiplexed pixel response to the anti-aliased and demultiplexed counterpart. In this paper, we propose to use a novel deep neural networks framework as the demultiplexer. By using demultiplexer and classic photometric stereo, our method can reconstruct a dense and accurate surface normal from a single-frame colored image without any prior information nor extra information injected. We build an imaging device to collect images of different materials under colored lights and white lights. We conducted extensive experiments on our data set and a public data set. The results show that the proposed fully connected network successfully demultiplexes the colorful image and produces satisfactory surface estimation.