Wednesday, 23 March 2011

Context-aware video resizing

I watched a demo of seam carving for video retargeting. Basically, they're cutting out the "most boring" bits of video frames to resize video in a context-aware way. It's not perfect, especially when there are big foreground objects, but it's a cool effect to see. The seams they were using were across all frames and intended to resize vertically or horizontally.

I wonder what would happen if you applied the seams in the time dimension, cutting out dynamic frames to "retarget" a video to a shorter time. There's nothing in the algorithm that would prevent doing so, and I think the results would be pretty interesting. It would probably come out looking a bit hyperactive. For instance, if two people were talking in turn, such an algorithm might cut out the parts where each person is not talking, compressing the conversation until the participants just talk over each other the whole time.

Mokalus of Borg

PS - In some movies, you might not notice the difference.
PPS - And some important scenes might go missing because they're too visually simple.

No comments: