Photometric Stereo Based on Multiple Kernel Learning.

Photometric stereo is a widely used surface reconstruction method which can estimate surface normals of an object from its images captured under different lighting conditions by a fixed camera. To deal with non-Lambertian reflections efficiently, kernel regression based photometric stereo has been proposed and achieved promising results. However, in practice, different data-sets often require different kernels, and the existing methods need selecting and tuning the predefined kernel manually. This is not user-friendly since it’s hard to find the best kernel for different data-sets. Furthermore, an improper kernel is very likely to degrade the performance. In this work, we adopt multiple kernel learning to handle this problem. The proposed method learns an optimal consensus kernel from multiple predefined kernels by automatically assigning the most suitable weights for different base kernels. The proposed method is tested on various data-sets, and the experiment results show that our multiple kernel based model outperforms the single kernel based method.