| Summary: | 碩士 === 大同大學 === 通訊工程研究所 === 93 === Depth from stereo has been one of the most actively researched topics in computer vision. In this work, we present an effective and efficient depth estimation algorithm suitable for implementation on a commodity PC. The proposed algorithm is based on the MML (Multiple Mipmap Levels) algorithm presented by Yang and Pollefeys[1]. Functionally, the MML method used correlation windows of fixed size for depth estimation. The main problem of correlation-based method is on the selection of window size. In areas of low texture, the window size cannot be too small, or it would not cover enough structure to resolve the disparity. On boundaries of distinct objects, large windows might cover objects of different depths which would result in false estimation.
We propose a method aiming at solving the problem of the MML algorithm by adaptively adjust mipmap levels based on the intensity variation of the image. In areas of high intensity variation, we use higher-resolution mipmap levels for the disparity estimation; while in areas of low intensity variation, we use lower-resolution mipmap levels. We compute gradient as an alternative representation of intensity variation due to the simplicity and efficiency of implementation.
The proposed algorithm shows that it is able to improve the disparity estimation in areas of low texture and on boundaries of distinct objects. We successfully demonstrate the system by processing several stereo images pairs and obtain satisfactory experimental results.
|
|---|
