Aerial-Guided Groupwise CRFs for Multiple Panorama Depth Estimation with Scale-Consistent

The task of depth estimation is important in many areas. However, current methods struggle with outdoor scenes that cover a wide range of depth scales, due to the oversimplification or strong prior of scale information. In this paper, we introduce the Aerial-Guided Groupwise Hypergraph CRFs framework for estimating depth in multiple panoramas with scale consistency. Specifically, we improve the estimation of depth across multiple panoramic images while keeping scales consistent. On one hand, we exploit scale constraints between aerial and ground from satellite perspectives and establish inter-correlations among ground images to maintain relative consistency. On the other hand, we propose a hierarchical cross-view aggregation module that facilitates relative position encoding and scale attention. Utilizing these cross-view correlations, we construct a groupwise hypergraph CRFs to optimize the scaling of depth. Additionally, we have developed a specialized dataset for multiple depth estimation, comprising satellite images that encompass all ground locations within a group. Extensive experiments and ablation studies on this dataset demonstrate the effectiveness of our method across various application scenarios.