Panoramic image depth estimation based on deformable convolution and attention mechanism

Currently, mainstream panoramic depth estimation methods primarily focus on correcting distortion effects. For Convolutional Neural Networks (CNN) based on standard convolution, it is difficult to fully perceive the complexity of the panoramic structure due to the fixation of its receptive fields. In this research, we employ deformable convolutions as a replacement for traditional standard convolutions, allowing sample points to adapt more flexibly to changes in object shapes, thereby extending the effective receptive field. Furthermore, this study incorporates the Bottleneck Attention Module (BAM) attention model in the feature fusion module to enhance focus on key areas.Through a series of experiments, the effectiveness of deformable convolution and BAM is verified.