Shape from shading using robust statistics

This paper describes how robust statistics can be applied to control the smoothness constraint in shape from shading. Conventionally, the smoothness-error is added to the data-closeness (or brightness-error) as a quadratic regularizer. The novelty of the work reported is to replace the quadratic regularizer with a robust error-kernel. Experiments on real and synthetic data reveal that robust regularizer prevents over-smoothing at high-curvature surface features and results in a significant improvement in reconstruction error