Depth Estimation based on a Single Close-up Image with Volumetric Annotations in the Wild: A Pilot Study
Frank P.-W. Lo, Yingnan Sun, Benny P. L. Lo · 2019
A novel depth estimation technique based on a single close-up image is proposed in this paper for better understanding of the geometry of an unknown scene. Previous works focus mainly on depth estimation from global view information. Our technique, which is designed based on a deep neural network framework, utilizes monocular color images with volumetric annotations to train a two-stage neural network to estimate the depth information from close-up images. RGBVOL, a database of RGB images with volumetric annotations, has also been constructed by our group to validate the proposed methodology. Compared to previous depth estimation techniques, our method improves the accuracy of depth estimation under the condition that global cues of the scene are not available due to viewing angle and distance constraints.