Pose-Depth Joint Optimization and Adaptive Sampling for Self-Supervised Multi-Frame Depth Estimation

Self-supervised multi-frame depth estimation encodes geometric information of different depths by constructing a Cost Volume and provides geometric priors for multi-frame depth estimation by minimizing the matching cost. However, existing methods overlook the potential impact of inter-view pose estimation errors and depth sampling ranges on the quality of geometric priors, which limits matching accuracy. In this paper, we propose an optimization strategy that combines the wide-baseline constraint in monocular depth estimation with the cyclic consistency constraint in visual odometry to alleviate the issue of cumulative errors between views and improve the matching quality of the multi-frame Cost Volume. Additionally, to address the limitations of traditional fixed-range sampling, this paper adaptively adjusts the sampling width of the Cost Volume based on the uncertainty of monocular depth estimation, reducing the interference from similarly textured regions. The experimental results show that on the KITTI dataset, our method achieves significant improvements: compared to ManyDepth and MoveDepth, our method improves the Abs Rel metric by 14.3% and 5.4%.