Uncertainty Guided Iterative Optimization Dense Matching Network for Occlusion Aerial Stereo Imagery

To address the challenge of local disparity estimation errors caused by large occlusions in aerial stereo imagery, an UGNet is proposed for dense matching. The network begins with a multi-scale cost aggregation-based disparity uncertainty estimation module, which generates an initial coarse disparity map while estimating uncertainty to effectively localize occlusion areas. Subsequently, an uncertainty-guided GRU-based disparity optimization module is introduced to iteratively reconstruct the disparity map, incorporating cross-correlation and both local and global contextual information. This approach mitigates error propagation in heavily occluded areas and improves estimation accuracy. Experimental results on the public WHUStereo and UAVStereo datasets demonstrate that the proposed method significantly outperforms existing dense matching networks in terms of EPE and 3PE, particularly in occlusion areas. The results indicate strong potential for accurately processing urban aerial stereo imagery under severe occlusion conditions.