Robust correspondence search under photometric variations and image ambiguity

Abstract

Establishing visual correspondences in multiple images is the core of many computer vision applications such as stereo, motion, tracking, object recognition, etc. To resolve the correspondence problem, many methods have been proposed for decades. However, it is yet to be solved in some aspects - it is still difficult due to photometric variations such as illuminant color changes and specular reflection and inherent image ambiguity. In this dissertation, we focused on the area-based correspondence search approach and proposed a few methods for area-based correspondence search under photometric variations and image ambiguity. This dissertation is composed of two parts - the first part is for analyzing and compensating photometric variations and the second part is for robust correspondence search under image ambiguity. In the first part, a few methods are proposed based on the dichromatic reflection model to recover the actual colors of scene surfaces - diffuse colors of scene surfaces under the canonical illumination - regardless of photometric variations such as illuminant color change and specular reflection. Recovering the actual colors of scene surfaces is important for the accurate raw matching cost computation under photometric variations. First of all, a voting-based method for specular pixel separation is proposed. In this method, specular pixels are rapidly separated by locally comparing the approximated diffuse colors of image pixels. Separated specular pixels are then used for illuminant chromaticity estimation. To estimate the illuminant color, a new dichromatic method based on the specularity-invariant dichromatic slope and the three-dimensional dichromatic lines space is presented. In this approach, illuminant color estimation is achieved by analyzing the dichromatic line space after projecting separated specular pixels into the dichromatic line space. Especially, for uniformly colored surfaces, illuminant color estimation and reflection components ...