Dense Depth Estimation with Absolute Scale

Abstract

Considering the difficulties in estimating depth from single image, in this paper, we propose a method to obtain the absolute scale depth map by combining the convolution neural network and depth filter. We compute relative transformation between consecutive frames by direct tracking features, which are extracted from RGB images and whose depthes are predicted by deep network, and then optimize relative motion by searching for a better feature alignment in epipolar line, and finally update every pixel depth of the reference frame by depth filter. We evaluate the proposed method on the open dataset comparison against the state of the art in depth estimation to evaluate our method.