Limitations
Feature Extraction and Tracking
 | Banding - Extracted points are in bands |
Banding
| Points which appear good in one image may or may not necessarily be good points to track in the other frames. |
| Points may get occluded in some frames. |
| Reflections off surfaces causes the extraction and tracking algorithms to perform incorrectly. |
Factorization
The factorization process depends greatly in the correctness of the measurement data. A few outliers can skew the results and hence producing an acurate shape. Also, due to data fitting and approximation computations, the shape coordinates don't come out as the real 3D coordinates of the feature points.
Meshing
Our mesher is currently limited to polygonal meshes that are convex. It works fine for most objects, including non-convex ones, by "jiggling" the vertices by a small random factor. However, in order to accomodate really non-convex polygons, the algorithm has to be changed.
3D View Synthesis
The textures currently come from the initial image that was used for tracking, for maximum accuracy. If our algorithms were less modular, one might be able to integrate the texture from ALL of the images, giving a higher degree of correctedness as the texture from the images will be suitably transformed for each of the differing viewpoints. Thus, one might for example use the correct texture map for one viewpoint and the nearest fit texture map for another view point, changing the texture on a per viewpoint basis.