Illumination insensitive methods for visual comparison
Image comparison is a fundamental component in many computer vision tasks such as recognition, alignment and tracking. Variation in lighting is critical to image comparison, because it dramatically affects the appearance of an object in an image. Current methods approach this problem by focusing on effects created by discontinuities in matte objects, which are insensitive to lighting changes. However this approach is very limited because it treats other effects as unmodeled noise, including the effects of smooth, untextured surfaces and the effects of shiny objects that produce significant highlights. These are present in most real objects. I show that these properties can be modeled, and they provide a rich source of information. I tackle the problem presented by smooth surfaces with no edges or texture by using a whitening tool from signal processing theory to design a superior measure of image comparison. This can provide a component in a general comparison method that also integrates past approaches. Recognition of shiny objects is also very challenging, since the appearance of the highlights they produce changes drastically with the viewing conditions. I show that using a simple qualitative model of specular reflection I can judge the consistency of specularities with 3D object geometry and use this consistency to identify very challenging transparent objects such as wine glasses. Next, I integrate this knowledge about the highlights with previous methods for matte objects. This allows recognition of glossy, smooth objects, such as pottery, which are very challenging for existing methods.