[nairoz]

I guess yes. Having worked on video processing, it's always better if you can stabilize because it significantly reduces the number of unique tokens, which would be even more useful for the present method. However, you probably lose in generalization performance and not all videos can be stabilized.