Regularized Local Linear Model with Core Neighbors for Reflectance Estimation

Regularized local linear model has been shown to be an effective approach for reflectance estimation. This approach estimates the reflectance of each test point by the linear combination of only its neighbors. The choice of neighbors is of crucial importance to achieve high estimation accuracy. We propose a principal components analysis based neighborhood selection method to reduce model bias. The idea is to find a subset of the test point's nearest neighbors, which we term core neighbors, that have the least reconstruction errors by retaining only the main principal components. Experimental results are provided to validate the effectiveness of the proposed approach.