| 摘要 |
The invention proposes a new motion layer extraction method using reference layer mosaics. One motivation is to work with distant images (whose motion in between is more discriminative) while reducing the number of pairs of images to be handled benefiting from the fact that layer mosaics represent the video sequence in a compact way. In this model, not only extracted layer supports for mosaicing are used but also layer mosaics for motion segmentation. In the proposed approach, instead of linking temporally many images pairs from the sequence, the currently processed original image is linked temporally with two sets of motion layer mosaics only. Thus a new image-to-mosaic data term replaces the usual image-to-image motion data term in the objective classification function. |
