Texture classification for material recognition

Summary form only given. In this presentation, classifying material texture from a single image under unknown viewing and lighting conditions will be discussed. Material recognition is important: an image may be considered as a collection of imaged materials. Humans can easily distinct the material properties of many surfaces. However, the appearance of materials changes significantly under different imaging settings, depending on the settings themselves and on the physical properties of the material. Understanding these aspects turns out to be crucial for material recognition. Visual appearance depends on the illuminant position and intensity, the incidental viewpoint of the camera, material properties of the objects, and the composition of the scene, resulting in occlusions and clutter. A natural starting point is to investigate the physical formation process of materials. That is, to study the physical laws of light interacting with material and reflecting towards the camera, this will allow appearance to be separated from intrinsic material properties. A sketch of the issues involved in developing models of the image formation process will be given, taking into account the variation in material appearance due to the illumination environment and camera geometry. Another major aspect is the composition of scenes as induced by the viewing geometry and the projection of the 3D world onto a 2D picture. Hence, I will consider the statistics in the scene as induced by the viewing geometry, which will explain another source of variation in material properties. Studying such statistical models allows to capture many of the essentials of material properties, taking the inherent correlations present in image data into account. The current and successful approach to this task is to treat it as a statistical learning problem and learn a classifier from a set of training images, but this requires a sufficient number and variety of training images. The modelling steps which wi- - ll be addressed capture the appearance variation of material images, allowing their properties to be summarized using parameterized representations of natural image statistics. After suitable invariant transformations, the parameters can be used to characterize the material in a codebook model. The number of training images required can be drastically reduced (to as few as three) by synthesizing additional training data using photometric stereo. Machine learning techniques can be devised to learn the appearance characteristics of various materials. The proposed methods will be demonstrated on the PhoTex collection and the Amsterdam Library of Textures (ALOT). ALOT is a color image collection of 250 rough textures, recorded for scientific purposes. In order to capture the sensory variation in object recordings, the viewing angle, illumination angle, and illumination color are systematically varied for each material. One hundred images of each material has been recorded, yielding 25,000 images for the collection. This collection is similar in spirit as the CURET collection. Although the number of view-illumination directions per material is only half the BRDF resolution of CURET, ALOT extents the number of materials almost by a factor 5, and it improves upon image resolution and color quality. Furthermore, different light source colors have been added to test (texture) color constancy algorithms.