Dynamosaicing: mosaicing of dynamic scenes

Area of Science:

• Computer Vision
• Digital Image Processing
• Video Editing

Background:

• Traditional video editing often struggles with precise temporal control of dynamic events.
• Aligning video frames for panoramic creation is challenging, especially with moving objects.
• Existing methods like 'brightness constancy' are insufficient for dynamic scene alignment.

Purpose of the Study:

• To develop methods for manipulating chronological time in video editing.
• To enable the creation of panoramic movies by aligning events in time.
• To address challenges in aligning dynamic scenes and avoiding visual artifacts.

Main Methods:

• Constructing an aligned space-time volume from input video.
• Sweeping a 2D time front through the space-time volume to generate new image sequences.
• Proposing 'dynamics constancy' for aligning dynamic scenes, surpassing 'brightness constancy'.
• Formulating optimal time front geometry as a minimal cut problem in a 4D graph, solved with max-flow methods.

Main Results:

• Successful manipulation of event timing, allowing for selective slowing and speeding of actions.
• Generation of panoramic movies by aligning events to a single time interval.
• Mitigation of visual seams and artifacts within moving objects through optimized time front geometry.
• Demonstration of the effectiveness of 'dynamics constancy' over 'brightness constancy' in dynamic scene alignment.

Conclusions:

• The proposed methods offer advanced control over temporal aspects in video editing.
• The 'dynamics constancy' principle and max-flow optimization effectively handle dynamic scenes and reduce artifacts.
• This work advances the creation of seamless panoramic videos with manipulated event timelines.