• Introduction
• Proposal
• Projective Geometry
• Registration
• Global Registration
• Composing
• Implementation and Results
• Conclusion
• References
• Appendix

Conclusion

Results and Limits

In this project we studied and implemented methods to register and merge multiple images into a high resolution spherical panorama. Our general framework (described in the introduction) is common to many of the papers we studied dealing with image mosaicing. The main differences between those mainly lie in the specific techniques used for the pair-wise registration, global registration and blending phases, as well as in the complexity of the camera model used. We intentionally chose relatively simple solutions from these available methods, which made it possible to implement a complete solution and to get interesting results within the available time. Nevertheless, it appears that this first order approach yielded sufficiently compelling results.

Phase Correlation and Feature-Based Correlation

We focused on the comparison of two different methods for pair-wise image registration :

• Phase-correlation proved to be fast and reliable. We used it for most of the panoramas presented in this report. However it appears to be limited to cases where the rotation and perspective distortions are small enough, that is to sequences of numerous narrow-angle shots. Moreover, phase-correlation needs a large amount of overlap (> approx. 30%) between the images to work correctly.
• On the contrary, the feature-based method is capable of handling any kind of projective transforms, so it can register wide-angle shots. It can find a rough registration even if the overlap is small. But it is likely to fail in the absence of distinctive features (as in sky images), and is sensitive to noise and moving objets. It also appeared to be significantly slower than phase-correlation (and more complex to implement).

Possible Improvements

Lack of time prevented us from doing the complete half-sphere panorama we initially planned to render (see the proposal), though our program should have had all the necessary elements to create one. We thought of some other improvements to increase the quality of the panoramas and make the use of the program easier. Notably an automatic extraction of the focal length from the image data would make the tedious task of approximating its value by trial and error unnecessary. The multi-resolution spline blending approach might have enabled us to get crisper images. Lastly a moving object detection heuristic like the one described in [Davis98] would have eliminated the "ghost artifacts" we got in some of our images.