Structure from Motion

Wiki Home ▸ Structure from Motion

Features and Tasks

The MVE SfM system is currently in construction. The code is available from the "sfm" branch. It currently features:

• SIFT and SURF feature detection and descriptor extraction
• Two-view feature matching using accellerated exhausive matching
• Fundamental matrix estimation using (at least) eight 2D-2D point correspondences
• Enforcing fundamental and essential constraints on a matrix
• Camera pose estimation from essential matrix
• Fundamental matrix computation from pose between two views
• Normalization of 2D and 3D points (for robust algebra)
• Triangulation of 2D-2D correspondences given pose using DLT transform
• Pose estimation using RANSAC for 2D-2D and 2D-3D correspondences

Things that are still to do:

• Initial pair selection
• Next view selection
• Bundle adjustment
• Proper data structures to add the next view
• etc.

Relevant Literature

• Multiple View Geometry in Computer Vision, 2nd Edition (Hartley & Zisserman, Book, 2003)
• SIFT: Distinctive Image Features from Scale-Invariant Keypoints (Lowe, IJCV 2004)
• SURF: Speeded Up Robust Features (Bay et al., CVIU 2008)
• Robust Rotation and Translation Estimation in Multiview Reconstruction (Martinec et al., CVPR 2007)
• Bundle Adjustment in the Large (Agarwal et al., ECCV 2010)
• Photo tourism: Exploring photo collections in 3D (Snavely et al., SIGGRAPH 2006)
• Modeling the World from Internet Photo Collections (Snavely et al., IJCV 2007)
• Skeletal graphs for efficient structure from motion (Snavely et al., CVPR 2008)
• Building Rome in a Day (Agarwal et al., ICCV 2009)