3D Reconstruction and Understanding of Indoor Scene Based on Single Image

Abstract

Given a photo of a room, through our automated system, try to reconstruct a three-dimensional scene similar to the photo. Among them, the indoor objects come from the objects in the ShapeNet model library, which enables our system to produce a high-quality output result. The method we use is to use a full convolutional neural network to train images and compare indicators, use Faster-RCNN to perform multitasking object recognition and target detection, and then iteratively optimize the position and proportion of objects in the room, so that the scene renders a three-dimensional map Best match with input photo.