[Contextcapture] Advancements in interior 3D modelling

Interesting paper in advancements for 3d reconstruction of interiors.

Panoramas,   if  equipped  with  the  depth  information,could  enable 

1. full  stereoscopic  VR  experiences; 
2. 3Dmodeling of surrounding environments; and
3. better scene understanding. 

However, panoramic 3D reconstruction has been  a  challenge  for  Computer  Vision  due  to  the  minimal  parallax,  which  is  important  to  reduce  stitching  artifacts but makes it difficult to utilize powerful multi-view reconstruction techniques.  The lack of texture exacerbates the situations for indoor scenes. The reconstruction accuracy of single-view methods is still far below the production level and successful panoramic 3D reconstruction has been demonstrated only with the use of special hardware such as a depth camera.

This  paper  proposes  a  novel  Structure  from  Motion(SfM)  algorithm  for  indoor  panoramic  image  streams  ac-quired by standard smart phones or tablets. The key idea is the fusion of single-view and multi-view recon-struction techniques. In the past, 3D vision community has rarely seen such fusion, mainly because single view methods are too “rough” to be directly used with the multi-view techniques. We seek to utilize single-view  techniques  to  effectively  detect  geometric  re-lationships of lines (e.g., detecting 2D lines as coplanar in3D), which in turn yield precise geometric constraints to be used in multi-view 3D reconstruction.

https://arxiv.org/pdf/1612.01256.pdf