Stereoscopic view synthesis with progressive structure reconstruction and scene constraints

Area of Science:

• Computer Vision
• Computer Graphics

Background:

• Depth image-based rendering (DIBR) is crucial for 2D-to-3D conversion, using texture and depth maps to create virtual views.
• Current DIBR systems face challenges with disocclusion holes, where existing inpainting methods struggle due to processing limitations and hole characteristics.

Purpose of the Study:

• To develop a novel learning-based method for stereoscopic view synthesis that effectively addresses disocclusion holes in DIBR.
• To improve the quality and accuracy of synthesized views in 2D-to-3D conversion applications.

Main Methods:

• Proposes a progressive structure reconstruction strategy to restore disocclusion regions, moving beyond traditional texture inpainting.
• Incorporates special scene cues as network constraints to mitigate hallucinated structure mixtures between different layers.

Main Results:

• The proposed method demonstrates significant improvements in handling disocclusion holes compared to existing approaches.
• Empirical evaluations validate the effectiveness of the progressive structure reconstruction strategy for stereoscopic synthesis.

Conclusions:

• The novel learning-based method is well-suited for stereoscopic view synthesis in 2D-to-3D conversion.
• The approach offers a more robust solution for disocclusion hole restoration in DIBR systems.