Quels prétraitements pour la stéréophotométrie non calibrée ?

Photometric stereo is a 3D-scanning technique which consists in taking several photographs of an object under the same angle but under different lightings. When the lightings are unknown, the problem is more complex. To solve it, further assumptions are required, amongst which the Lambertian assumption and the constraint of integrability. We show that some well-chosen preprocessings on the images may limit the influence of possible deviations from those assumptions. We suggest that the suspected pixels should be discarded. We validate these improvements on a database of real images of faces available on the web.