A population Monte Carlo approach to estimating parametric bidirectional reflectance distribution functions through Markov random field parameter estimation
Doctoral Thesis
2009
Permanent link to this Item
Authors
Journal Title
Link to Journal
Journal ISSN
Volume Title
Publisher
Publisher
University of Cape Town
Department
License
Series
Abstract
In this thesis, we propose a method for estimating the parameters of a parametric bidirectional reflectance distribution function (BRDF) for an object surface. The method uses a novel Markov Random Field (MRF) formulation on triplets of corner vertex nodes to model the probability of sets of reflectance parameters for arbitrary reflectance models, given probabilistic surface geometry, camera, illumination, and reflectance image information. In this way, the BRDF parameter estimation problem is cast as a MRF parameter estimation problem. We also present a novel method for estimating the MRF parameters, which uses Population Monte Carlo (PMC) sampling to yield a posterior distribution over the parameters of the BRDF. This PMC based method for estimating the posterior distribution on MRF parameters is compared, using synthetic data, to other parameter estimation methods based on Markov Chain Monte Carlo (MCMC) and Levenberg-Marquardt nonlinear minimization, where it is found to have better results for convergence to the known correct synthetic data parameter sets than the MCMC based methods, and similar convergence results to the LM method. The posterior distributions on the parametric BRDFs for real surfaces, which are represented as evolved sample sets calculated using a Population Monte Carlo algorithm, can be used as features in other high-level vision material or surface classification methods. A variety of probabilistic distances between these features, including the Kullback-Leibler divergence, the Bhattacharyya distance and the Patrick-Fisher distance is used to test the classifiability of the materials, using the PMC evolved sample sets as features. In our experiments on real data, which comprises 48 material surfaces belonging to 12 classes of material, classification errors are counted by comparing the 1-nearest-neighbour classification results to the known (manually specified) material classes. Other classification error statistics such as WNN (worst nearest neighbour) are also calculated. The symmetric Kullback-Leibler divergence, used as a distance measure between the PMC developed sample sets, is the distance measure which gives the best classification results on the real data, when using the 1-nearest neighbour classification method. It is also found that the sets of samples representing the posterior distributions over the MRF parameter spaces are better features for material surface classification than the optimal MRF parameters returned by multiple-seed Levenberg-Marquardt minimization algorithms, which are configured to find the same MRF parameters. The classifiability of the materials is also better when using the entire evolved sample sets (calculated by PMC) as classification features than it is when using only the maximum a-posteriori sample from the PMC evolved sample sets as the feature for each material. It is therefore possible to calculate usable parametric BRDF features for surface classification, using our method.
Description
Keywords
Reference:
Louw, M. 2009. A population Monte Carlo approach to estimating parametric bidirectional reflectance distribution functions through Markov random field parameter estimation. University of Cape Town.