Iterative median algorithm based on a cost cube for stereo matching

Wide featureless and highly repetitive regions have always been critical challenges in stereo matching. Most existing cost measurement methods cannot effectively extract decisive features for pixels located in these special regions, which leads to an increase in misalignment rates. In this paper, we propose a feature extraction method based on arms. It reflects the location information of pixels through the multi-directional extension of the arms. To prevent the arms from stopping prematurely due to fake boundaries, we set dynamic thresholds to ensure the reliability of the arm lengths. The length of the feature arms, as a part of the cost volume, is used to quantitatively measure the similarity. Comparison experiments show that the introduction of feature arms significantly improves the matching accuracy of these special regions. In addition, we propose a new cost optimization strategy, to our knowledge, called iterative radiation optimization with self-checking mechanism. Unlike traditional cost aggregation, it abandons the concept of paths and instead adopts a scheme where valid disparities radiate to the invalid ones. The validity of pixels is determined by the self-checking mechanism. Through continuous self-checking and iterative radiation optimization, the error rate can be reduced to a minimum. Qualitative and quantitative experiments on the Middlebury dataset validate the excellent matching accuracy of our algorithm. It has great potential for widespread extension in applications such as augmented reality, autonomous driving, and 3D reconstruction.