The paper proposes MotionMaster, a training-free camera motion transfer model that can effectively decouple camera motion and object motion in videos.
The key highlights are:
One-shot Camera Motion Disentanglement: MotionMaster can extract camera motion from a single source video by separating the moving objects and background regions, and estimating the camera motion in the foreground region based on the motion in the background.
Few-shot Camera Motion Disentanglement: To handle complex and diverse object motions, MotionMaster proposes a window-based clustering technique to extract the common camera motion from multiple videos with similar camera motions.
Camera Motion Combination: MotionMaster investigates the additivity and positional composition ability of camera motions, and proposes a camera motion combination method to achieve flexible camera control, enabling the combination of different camera motions and application of different camera motions in different regions.
Extensive experiments demonstrate the superior performance of MotionMaster in both one-shot and few-shot camera motion transfer, significantly improving the controllability and flexibility of camera motions compared to existing methods.
To Another Language
from source content
arxiv.org
Önemli Bilgiler Şuradan Elde Edildi
by Teng Hu,Jian... : arxiv.org 04-25-2024
https://arxiv.org/pdf/2404.15789.pdfDaha Derin Sorular