With the film and television industry’s increasing demand for stereoscopic visual effects and immersiveness, existing depth estimation methods are prone to errors in complex lighting, fast motion, and weakly textured areas, resulting in depth discontinuities and artifacts in stereo images. To address this issue, this paper introduces a depth estimation method for stereo photography based on multi-view visual matching. Combining camera geometry modeling with high-precision calibration techniques, this method constructs a multi-baseline camera array to enhance depth perception stability. The algorithm employs a strategy combining multi-scale feature matching with semi-global optimization, improves depth consistency through graph cuts and belief propagation models, and introduces hole filling and guided filtering in the post-processing stage to preserve object edge details and visual continuity. Experimental results show that the introduction of bilateral filtering reduces the MAE to 0.85 px, the bad pixel rate to 7.2%, the edge preservation index (EPI) to 0.81, and the hole rate to 4.7%, demonstrating that filtering significantly reduces noise and preserves edges. Further employing guided filtering reduces the MAE and bad pixel rate to 0.79 px and 6.8%, respectively, and improves the EPI to 0.85, demonstrating that guided filtering outperforms bilateral filtering in preserving edge details.
Jincheng Xiao (Thu,) studied this question.