In this study, we investigate a depth recovery method based on optical flow from two consecutive frames with relative motion between the object and the camera. Multi-resolution processing is suitable for high-density depth recovery that avoids aliasing and preserves discontinuities. Propagation of the recovery results from the low-resolution layer to the high-resolution layer is an important issue. In this paper, we propose a method based on variational Bayesian inference. By computing the posterior distributions of the depth and motion parameters at each layer using the mean-field approximation and converting them into the prior distribution of the upper layer, it is possible to propagate the depth and motion information simultaneously. The effectiveness of the proposed method was quantitatively evaluated using artificial images, and its practicality of the system was also confirmed qualitatively.