Panoramic images,with their wide field of view and comprehensive scene information,including spatial structure and details,provide accurate and detailed depth information and have gained significant attention in the fields of autonomous driving and robot navigation.Traditional panoramic image projection methods,such as Equirectangular Projection(ERP),offer a rich scene view but suffer from distortion and the inability to capture complete contextual information,while Cubemap Projection(CMP)avoids distortion,it has a limited field of view and discontinuous boundaries.To address these issues,we propose a dual-branch panoramic image depth estimation method called DeCSBAFuse,combining equirectangular distortion aware convolution and a hybrid attention mechanism.The key innovations of the proposed method are as follows:introducing an ERP branch based on EDPC,integrating vertical image information to fully preserve contextual information;designing a CMP branch based on channel and spatial attention mechanisms,incorporating batch attention to better handle boundary issues between different faces of CMP.Compared to mainstream panoramic image depth estimation methods,DeCSBAFuse demonstrates more accurate depth prediction results on three benchmark datasets,especially in areas with detailed textures,where it can predict clearer depth boundaries.
维普
万方数据