Guided Robust Matte-Model Fitting for Accelerating Multi-light Reflectance Processing Techniques

Abstract

The generation of a basic matte model is at the core of many multi-light reflectance processing approaches, such as Photometric Stereo or Reflectance Transformation Imaging. To recover information on objects’ shape and appearance, the matte model is used directly or combined with specialized methods for modeling high-frequency behaviors. Multivariate robust regression offers a general solution to reliably extract the matte component when source data is heavily contaminated by shadows, inter-reflections, specularity, or noise. However, robust multivariate modeling is usually very slow. In this paper, we accelerate robust fitting by drastically reducing the number of tested candidate solutions using a guided approach. Our method propagates already known solutions to nearby pixels using a similarity-driven flood-fill strategy, and exploits this knowledge to order possible candidate solutions and to determine convergence conditions. The method has been tested on objects with a variety of reflectance behaviors, showing state-of-the-art accuracy with respect to current solutions, and a significant speed-up without accuracy reduction with respect to multivariate robust regression.