Dense stereo processing requires a critical step that called cost aggregation or cost optimization. Most of the cost aggregation methods are evaluated on close range stereo images from Middlebury or KITTI datasets. While the effect of cost aggregation on high resolution satellite stereo processing has not yet been sufficiently evaluated. In this paper, three typical cost aggregation methods together with another approach which is a combination of these methods are evaluated on high resolution satellite stereo images and then are compared with LiDAR ground truth. These methods including Semi-Global Matching (SGM), Guided Filtering (GF), iterative GF (IGF), and SGM followed by GF (SGM-GF) with Census and Zero Normalized Cross Correlation (ZNCC) cost functions. Although the Census cost function has a good performance on the border of the objects and low blurring effects, the results of both cost functions, i.e. Census and ZNCC, have same treatment on all stereo methods. Also, in order to make an impartial assessment, for all stereo methods, we do not perform any disparity map refinement. The bad-pixel criteria with an absolute difference height error greater than 2 meters for SGM, GF, IGF, and SGM-GF methods is 36.7%, 34.8%, 33.8%, and 28.6% respectively. Also, the Normalized Median Absolute Difference (NMAD) error for SGM, GF, IGF, and SGM-GF is 1.29, 1.15, 1.06, and 0.94 meters, respectively. Overall, the experimental results on WV III stereo images demonstrate that the SGM method has lower accuracy and SGM-GF method is accurate than other methods.
展开▼
