Estimating Illumination Direction Of Three-dimensional Surface Textures Using Bag-of-features Algorithm

Texture is one of the ubiquitous surface features in natural scene, which producessurface appearance. As an important texture type,3D (three-dimensional) surfacetexture, which has rich surface structure information, is sensitive to illuminationvariance and viewing angle changes. Traditional texture analysis methods may notwork well on this kind of texture. This thesis reports a method of estimatingillumination direction of3D surface textures.In this thesis, a3D texture classification method using Bag-of-features algorithmwhich incorporates shadow and edge information is proposed. Textures wererepresented by Haar Active Basis model for histogram matching. BOF(Bag-of-features) algorithm is performed to training textures with known illuminationdirection. A predicted classification of test texture is first performed using SVM, theHaar Active Basis model is then applied on both test textures and training textures.The basis histograms are used for further matching. The illumination direction of thetest texture is assigned to the matched training texture.In the experiment, the proposed method was tested separately in two sections:1)the class of test texture exists in the training set;2) training set and test set have nooverlaps. The results showed that our proposed method effectively estimatedillumination directions of surface textures. Finally, we discussed failed cases forfuture work.