Confidence-guided stereo matching method with disparity space refinement for 3D reconstruction from optical remote sensing images

To address the susceptibility of stereo matching methods to errors in weak textures and disparity discontinuities, this paper proposes a confidence-guided stereo matching method with disparity space refinement for 3D reconstruction from optical remote sensing images. First, the initial disparity and matching cost are obtained by constructing the stereo matching cost. Then, multiple confidence features are jointly utilized and contextual information is combined to enhance both the confidence features and the estimated confidence. The constructed confidence is employed to refine the disparity search space and modulate the initial matching cost. A disparity space refinement strategy is proposed to assign each superpixel an independent disparity space based on the confidence and disparity relationships with its neighboring superpixels. Simultaneously, the initial matching cost is modulated by the constructed confidence, where high-confidence superpixels maintain similar matching costs, while low-confidence superpixels undergo cost smoothing. Finally, the refined disparity space and modulated cost volume are fed into an object-based iterative optimization process, with periodic updates to the disparity space to improve matching accuracy. Experimental results demonstrate that the proposed confidence-guided stereo matching method effectively improves the quality of 3D reconstruction, achieving an RMSE of 2.79 m, MAE of 1.68 m, MHE of 0.85 m, and COMP of 55.47% across multiple datasets, outperforming several mainstream 3D reconstruction methods.