Towards Robust and Physically Plausible Shaded Stereoscopic Segmentation

Dejun Wang (1) , Emmanuel Prados (2, 3) , Stefano Soatto (1)

1 UCLA-CS - Computer Science Department [UCLA]
2 UCLA Vision Lab
3 MOVI - Modeling, localization, recognition and interpretation in computer vision

We address the multi-view shape from shading problem, that is the recovery of 3-D shape, lighting configuration and surface albedo from multiple calibrated views. Previous approaches to this problem relied on physically impossible illumination models (negative light) and only work on constant albedo and resulted in biased estimates of shape and lighting positions. Furthermore, since the solution involves infinite-dimensional optimization, existing approaches were quite slow. We develop a new model that explicitly enforces positivity in the light sources with the assumption that the object is Lambertian and its albedo is piecewise constant and show that the new model significantly improves the accuracy and robustness relative to existing approaches. Furthermore, we show that the most computationally expensive step in the optimization can actually be solved in closed form. This significantly improves speed of convergence over existing schemes. We illustrate the behavior of our algorithm directly on the same data used by previous authors, so direct comparison is possible.

Domaines: Vision par ordinateur et reconnaissance de formes [cs.CV]

Dejun Wang, Emmanuel Prados, Stefano Soatto. Towards Robust and Physically Plausible Shaded Stereoscopic Segmentation. Computer Vision and Pattern Recognition Workshop (CVPRW), Jun 2006, New York, United States. ⟨10.1109/CVPRW.2006.204⟩. ⟨inria-00377430⟩