基于全局特征增强和几何先验的多视图立体方法OA

Although deep learning-based multi-view stereo(MVS)methods have made significant progress,the depth estimation accuracy of MVS for points in weak-texture and texture-less regions still requires further improvement.To address this issue,this paper proposes a multi-view stereo method based on global feature enhancement and geometric priors.The proposed method employs a global feature enhancement module to fuse raw RGB images with depth features while leveraging the global modeling capability of Transformers and the local detail extraction advantages of convolutional networks.This integration generates optimized multi-scale features,significantly enhancing the model's perception of weak-texture regions.Additionally,a geometric prior-guided module is introduced,which adopts a cross-stage cost volume fusion strategy.This strategy incorporates geometric information from the coarse stage into the fine-stage cost volume optimization process while utilizing convolutional networks for joint reasoning of geometric priors across different stages of the cascaded structure,consequently,improves depth estimation accuracy.Experimental results on multiple public datasets demonstrate that the proposed method outperforms existing state-of-the-art approaches,highlighting its effectiveness and superiority.