| Abstract: |  Recent work in Structure from Motion (SfM) has successfully built 3D models from unordered collections of images. Inspired by their success, we choose one of these as the baseline [55] for implementing our own 3D reconstruction pipeline. We introduce three improvements dealing with detection of a sufficient number of features for high resolution images, speeding up a standard RANSAC for epipolar geometry estimation and focal length estimation for sets of image files with insufficient information about cameras. We show that our pipeline performs just as well as, and in some cases better than, the baseline. To improve reconstruction of indoor scenes, we observe that traditionally employed features do not work well for significant appearance changes of local patches which is typical for large camera transformations. We propose to firstly understand the indoor scene as a whole and then exploit this knowledge to improve image matching. Inspired by recent success of monocular indoor scene reconstruction, we estimate a box-like scene model for every input image and rectify individual faces which are then utilized for matching. We show that using these additional matches brings a dramatic improvement in reconstructing challenging indoor scenes from images.
|
|---|