High Level 3D Structure Extraction from a Single Image Using a CNN-Based Approach.

摘要: High-Level Structure (HLS) extraction in a set of images consists recognizing 3D elements with useful information to the user or application. There are several approaches HLS extraction. However, most these based on processing two more captured from different camera views data form point clouds extracted images. In contrast and motivated by extensive work developed for problem depth estimation single image, where parallax constraints not required, this work, we propose novel methodology towards image promising results. For that, our method has four steps. First, use CNN predict image. Second, region-wise analysis refine estimates. Third, introduce graph segment semantic orientations aiming at identifying potential HLS. Finally, sections provided new architecture that predicts shape cubes rectangular parallelepipeds.