Analysis of multichannel texture representations from video mosaics

Video mosaics have been recently suggested as alternatives to key-frames for indexing video sequences in large video databases. We demonstrate that mosaicing techniques affect the textural content of the original sequence. Consequently the annotations derived during indexing may also be affected. In particular the high frequency content of the representation may not only reflect the high frequency content of the original sequence but also vertical stitches generated by the mosaicing process. Our experimental results also indicate that the processing used to combine the overlapping images (e.g. averaging) tends to destroy textural information. Eventually, we suggest that mosaics constructed without averaging, but from a subsampled sequence are a better alternative for video annotations. As a result we also stress the importance of the accurate registration and alignment of the images