| Home | People | Research | Publications | Demos |
| News | Jobs |
Prospective
Students |
About | Internal |
|
|
|
One of the most popular methods for summarizing video is through the construction of mosaics. The images in the sequence are registered and combined into an image mosaic that summarizes the content of the whole sequence. Traditionally, mosaics are constructed by measuring the motion between pairs of consecutive frames and then integrating the resulting pair-wise motion trajectories in order to find the mosaic location onto which each pixel in the sequence should be mapped. This fails to explore the consistency, over time, of the motion, and can lead to poor motion estimates in frame-pairs where an object is occluded or subject to variations of lighting. Poor motion estimates can then derail the integration process and lead to mosaics of very poor quality. This work explores the use of spatio-temporal motion constraints to derive mosaicking algorithms that are significantly more robust. The motion constraints and resulting algorithms algorithms are described in the CVPR 98 paper [ps,pdf]. Here we present some illustrative results of the benefits that can be obtained. In each case, we show a few frames of the original sequence, the mosaic obtained with traditional integration of pair-wise estimates, and the one produced by the spatio-temporal motion model. All mosaic construction parameters are the same, except the motion constraints enforced. |
| Sequence: |
|
|
|
|
| Pair-wise estimates: |
|
|
| Spatio-temporal estimates: |
|
|
| Sequence: |
|
|
|
|
| Pair-wise estimates: |
|
|
| Spatio-temporal estimates: |
|
|
| Sequence: |
|
|
|
|
| Pair-wise estimates: |
|
|
| Spatio-temporal estimates: |
|
|
©
SVCL
