Distortion-Aware Bi-Projection Fusion for 360∘ Monocular Depth Estimation via Coordinate Attention

Monocular depth estimation for 360∘ omnidirectional images is essential for immersive scene understanding and 3D reconstruction but remains challenging due to the non-uniform geometric distortions introduced by equirectangular projection (ERP). In particular, ERP suffers from latitude-dependent sampling bias, where the effective receptive field of standard convolution varies with the spherical latitude, leading to inconsistent feature representation and degraded depth prediction in high-distortion regions. To address this problem, this paper proposes a distortion-aware bi-projection fusion framework that integrates the global contextual continuity of ERP with the locally distortion-reduced geometric representation of cube map projection (CMP). The core component of the proposed framework is a Multi-scale Coordinate-Transformer Fusion (MCTF) module, which combines convolutional feature mixing, Transformer-based global context modeling, and Coordinate Attention-based spatial recalibration. By explicitly encoding vertical coordinate information into the fusion process, MCTF adaptively recalibrates feature responses according to latitude-dependent distortion patterns. Extensive experiments on the 3D60, Matterport3D, and Stanford2D3D benchmarks demonstrate that the proposed method consistently outperforms state-of-the-art omnidirectional depth estimation approaches. On the 3D60 dataset, our method reduces RMSE by 21.2% compared with the ERP baseline and achieves a 30% RMSE reduction in polar regions, where ERP distortion is most severe. These results validate the effectiveness of coordinate-aware feature calibration for robust 360∘ monocular depth estimation.