Enhanced Photometric Stereo with Multispectral Images
Tsuyoshi Takatani, Yasuyuki Matsushita, Stephen Ching-Feng Lin, Yasuhiro Mukaigawa, Yasushi Yagi · 2013
We introduce a technique based on multispectral images aimed at improving Lambertian photometric stereo. Many photometric stereo algorithms assume Lambertian reflectance, but deviations from this ideal produce errors in shape reconstruction. To alleviate this problem, we exploit the wavelength-dependence of material reflectance. Based on the observation that re-flectance at certain wavelengths for a given object is more Lambertian than at others, we propose a method for identifying such wavelengths by a matrix rank anal-ysis, and use them to achieve more accurate photo-metric stereo. We merge reconstruction results from different wavelengths to produce the final surface nor-mal map. Experimental results on synthetic and real data demonstrate the greater accuracy of this method compared to conventional photometric stereo based on brightness images. 1