3D Reconstruction of Single Indoor Image Based on Relational Network

To address the problems of object scale discrepancies and insufficient detail precision in reconstructing indoor scenes using only one image in 3D, a method for 3D reconstruction that increases the proportion of geometric features is proposed. The method for 3D reconstruction is based on scene understanding and uses the spatial layout estimation module to obtain the layout information of the scene. The method adopts the relational network to allocate the relationships between objects and increases the weight of geometric features to deal with the problem of mismatches in target scales in the reconstruction based on a single image. The reconstructed surface of the object is calculated using the sign distance function, and the surface details of the resulting reconstruction are optimized by combining them with a Marching Cubes algorithm. The method has been shown to be capable of obtaining a reasonable-scale reconstructed object model and the detailed message of the model, as evidenced by the experimental results.