Video texture synthesis with multi-frame LBP-TOP and diffeomorphic growth model

Area of Science:

• Computer Vision
• Computer Graphics
• Dynamic Texture Synthesis

Background:

• Video texture synthesis generates continuous, infinitely varying frame streams, crucial for computer vision and graphics.
• Current methods face challenges in producing high-quality synthesis results due to issues with frame representation and blending artifacts.

Purpose of the Study:

• To improve video texture synthesis performance by addressing frame representation and blending artifacts.
• To develop a novel approach for generating high-quality, continuous video streams.

Main Methods:

• A two-stage approach: video stitching and transition smoothing.
• Developed a spatial-temporal descriptor for effective frame representation, capturing spatial and temporal information.
• Employed a diffeomorphic growth model to suppress artifacts and enhance temporal continuity.

Main Results:

• The proposed method effectively captures spatial and temporal information for dynamic textures.
• Significantly suppressed artifacts, leading to improved synthesis quality and temporal continuity.
• Demonstrated robust performance on public databases and internet videos through qualitative and quantitative evaluations.

Conclusions:

• The novel video texture synthesis approach enhances frame representation and artifact suppression.
• Achieved high-quality synthesis results with improved temporal continuity.
• The method offers a significant advancement in generating dynamic textures for computer vision and graphics applications.