Bi-projection Fusion for 360 Depth Estimation Based on Coordinate Attention

Abstract

Depth estimation is a fundamental task in computer vision, essential for 3D scene reconstruction and spatial comprehension. While monocular depth estimation offers a cost-effective and flexible solution, its application to panoramic imagery is hindered by significant geometric distortions inherent in Equi-Rectangular Projection (ERP). To address this issue, we propose a panoramic depth estimation model based on bi-projection fusion. By leveraging the distortion-free characteristics of Cubic Projection (CMP), the proposed method compensates for structural deformations and reduces the impact of polar distortions in ERP representations. Central to our approach is a Multi-scale Coordinate-Transformer Fusion (MCTF) module, which integrates Transformer-based global context modeling with Coordinate Attention-based feature recalibration to effectively fuse ERP and CMP features. The MCTF module adaptively enhances informative spatial and channel-wise responses while suppressing redundant or distortion-sensitive features, thereby improving the geometric consistency of panoramic depth prediction. Experimental results on mainstream panoramic datasets demonstrate that the proposed bi-projection strategy significantly improves accuracy over single-projection methods, and the incorporation of the MCTF module further enhances depth estimation performance.