Video Mosaics - Results

Team Members

Shahzad Malik

Pablo Sala

Report

This page shows the results from the Video Mosaics project. For a more thorough discussion of the system, please download the full report which is available in PDF format.

Download Full Report [PDF]

Results

Cylindrical Mosaics

Room Sequence
** Click here to view as a panorama **

This cylindrical mosaic was generated from 110 input images that were captured with a slowly panning camera. The focal length was determined using the 8-parameter alignment algorithm (with a manual initial estimation via four corresponding points), after which the automatic coarse-to-fine Lucas-Kanade algorithm was initiated. Notice the gradual downward "drift" in the mosaic, since the rotational motion is not purely panning. Nevertheless, the results are quite good, except there is no gap-closing at the end of the mosaics due to time constraints. In fact, there are some images that are repeated at the both the start and ends of the sequence. The most interesting aspect of this mosaic was that it was generated without using a tripod, yet the simple translational alignment technique is still successful. This mosaic was generated in about 15 to 20 minutes on a P2-500 Mhz PC, with each input image having a resolution of 512x256x24.

Outdoor Sequence
** Click here to view as a panorama **

This is another cylindrical mosaic that was generated using a focal length that was precomputed with the 8-parameter algorithm. Once the focal length was computed, the automatic coarse-to-fine Lucas-Kanade algorithm was started. While the stitching algorithm aligned the images quite well, there are very noticeable bands of colour throughout the mosaic. This is a result of the input images not being captured with the gain-control locked. As a result, different snapshots of the scene were captured with varying intensities. Nevertheless, the alignment algorithm still worked quite well. A feathering algorithm, as described in the Szeliski and Shum technical report, could help reduce the effects of the noticeable intensity gradients, but due to time constraints we didn't get a chance to implement this. Interestingly, this entire sequence only consisted of 19 input images, unlike the previous sequence of 110. The coarse-to-fine alignment technique is largely responsible for this sequence being successfully generated. (This image sequence was downloaded from the website for a similar assignment at the University of Washington).

Planar Mosaics

Tower and Computer Desk

The tower scene on the left consists of two frames that were manually aligned by picking corresponding points in each image, and then refining the estimate using the 8-parameter algorithm. The computer scene on the right was made in a similar fashion. In both cases, the results are quite good. After alignment, the final images were projected onto a planar manifold. (Both of the original input images for each of the sequences were downloaded from the website of a similar assignment at another university.)