Analysis of Patch Shapes in Dense Matching for Reducing Disparity Errors in UAV Stereo Images Hongjin Kim (a), Chaeyeon Lee (b), Taejung Kim (b*)
a) Program in Smart City Engineering, Inha University
100 Inha-ro, Incheon 22212, Republic of Korea
b) Department of Geoinformatic Engineering, Inha University
100 Inha-ro, Incheon 22212, Republic of Korea
*tezid[at]inha.ac.kr
Abstract
Disparity estimation using stereo images acquired from unmanned aerial vehicle (UAV) becomes a core technology for constructing precise 3D spatial information, digital twins, and smart cities. However, dense matching based on fixed-size square-shaped patches often results in mismatches in high-frequency regions such as building boundaries. This is because the traditional square-shaped patches do not sufficiently account for the directional structure or depth discontinuities at object boundaries, and because they treat different regions within the patch equally for computing matching costs. As a result, information near the boundaries becomes averaged, leading to blurred edges and disparity diffusion. This phenomenon degrades the quality of 3D data and causes information loss in object boundary areas. To alleviate this phenomenon, this study analyzes disparity variations within stereo image pairs by applying various types of patches during the matching cost computation stage. The matching cost is calculated using zero-mean normalized cross-correlation (ZNCC).
Disparity maps were generated for each patch shape at multiple scales, and the characteristics of the resulting maps such as structural preservation at edges, information averaging, and disparity spreading were compared. Based on the analysis, a final disparity map was generated by combining multi-scale and multi-shaped patches. The effects of this fusion approach were evaluated on reducing mismatches near boundaries, suppressing disparity blurring, and improving the overall resolution and quality of the disparity map. Experimental results showed that the proposed method reduced information loss around object boundaries and improved the overall quality of the disparity map. This suggested that the proposed method not only effectively mitigated the boundary diffusion problem observed in conventional fixed-patch approaches but also contributed to the precision and practicality of UAV based 3D spatial information generation.
Keywords: Dense stereo matching, UAV image, Disparity estimation, 3D reconstruction
