3 D Reconstruction of Antique Stereo Pair Images Indriyati Atmosukarto, Anna Cavender, Chandrika Jayant
Objective • What is the best way to reconstruct 3 D scenes from antique stereo pair images? • How do we handle over-segmentation in more complicated stereo pair images?
Related Work • Single View Metrology [Criminisi et al ’ 99] – Uses only one image – Manual input required • Multiway Graph Cut [Boykov, et al ’ 98] – Min. Cut Algorithm Approximation with MRFs • Slanted Surfaces [Birchfield et al ’ 99] – Uses Boykov and affine parameters – Connected components
Our Work 1) Implement SVM 2) Implement Birchfield with Affine 3) Implement Graph Cut with User Constraints 4) Compare Results
Single View Metrology, Method, Modeling • • Compute 3 vanishing points (x, y, z) Compute homography Define planes for 3 D models Texture mapping to planes
Birchfield- Affine • Graph cut code supplied • Iteration method in paper • Our idea: Affine parameter estimation using feature matching (SIFT)
Graph Cut- User Constraints • Use Graph Cut • Run Connected Components • Use user input to constrain regions and deal with oversegmentation • Used lines and rectangles– Still very difficult even with user input!
Comparison • Disparity 3 d model • 3 d model disparity • Slanted surfaces
Results
Tsukuba Stereo
SVM
Graph Cut-User Constraints Graph Cut Ours
3 D Reconstruction of Disparity Map
Venus Stereo: Graph Cut User Constraints
3 D Reconstruction
Brussels Stereo- SVM
Graph Cut- User Constraints Graph Cut Line Constraint Rectangle Constraint
House Stereo
SVM
Graph Cut -User Constraints Graph Cut Ground Constraint More Constraints
Conclusion • Antique photos hard to work with! – Complex – Noisy – Don’t know camera parameters • Hard to compare methods • Segmentation techniques must be improved
Future Plans • Finish Affine Parameter Estimation • Automated Rectangles for Oversegmentation Problem -First requires better user input method! • Noise reduction for antique photos