Description
Components of the Project
- Use relevant tools (e.g. vl feat: http://www.vlfeat.org) to find interest points and correspondence (matches) between consecutive images.
Note: You can get 1 bonus point if you submit a workable implementation of your own for Harris corner detection and feature matching. (Lab assignment 2).
- Apply normalized 8-point RANSAC algorithm to find best matches. (Lab assignment 3+5)
- Chaining: Create point-view matrix to represent point correspondences for different camera views (Lab assignment 6).
- Stiching:
- Take blocks of the point-view matrix that are composed of three and four consecutive images.
1
- Estimate 3D coordinates of each block using TomasiKanade factorization (Lab assignment 4).
- By an iterative manner, stitch each 3D point set to the main view using the point correspondences i.e., finding optimal transformation between shared points in your 3D point clouds. Transformation between different sets can be found using Procrustes analysis. The MATLAB procrustes function can be used in the project.
- Eliminate affine ambiguity (consult your lecture slides on affine structure from motion).
- 3D Model Plotting: When you have the 3D point cloud of the castle, use the built-in surf function for the 3D surface plot. Then include RGB (texture) colour of the related points on surf visualization (interpolate colour values for the filled areas on the surface using the known points).
