StabStitch++: Unsupervised Online Video Stitching With Spatiotemporal Bidirectional Warps.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Video stitching often introduces warping shakes, degrading visual quality even with stable input videos.
• Existing methods struggle to balance spatial alignment and temporal stabilization effectively.

Purpose of the Study:

• To introduce StabStitch++, a novel framework for simultaneous spatial stitching and temporal stabilization in videos.
• To address the issue of warping shakes in video stitching using unsupervised learning.

Main Methods:

• Proposes a virtual midplane projection and a differentiable bidirectional decomposition module for spatial warping.
• Derives mathematical expressions for stitching trajectories by integrating spatial and temporal warps.
• Introduces a warp smoothing model with a hybrid loss for content alignment, trajectory smoothness, and online collaboration.

Main Results:

• StabStitch++ optimizes spatial alignment and temporal stabilization simultaneously, unlike previous methods.
• Achieves superior stitching performance, robustness, and efficiency compared to existing solutions.
• Enables a real-time online video stitching system.

Conclusions:

• StabStitch++ offers a significant advancement in video stitching by resolving warping shakes.
• The framework provides a no-compromise solution for both alignment and stabilization.
• The developed dataset and real-time system facilitate further research and application.