Deep Sparse-to-Dense Inbetweening for Multi-View Light Fields.

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• Light field (LF) imaging captures intensity and directional light information, surpassing traditional methods.
• Existing LF view synthesis struggles with sparse inputs and single-view robustness.
• Sparse-to-dense inbetweening addresses these limitations by generating dense views from sparse LF data.

Purpose of the Study:

• Introduce and define the sparse-to-dense inbetweening task for LF imaging.
• Develop a robust method for generating dense novel views from sparse multi-view LFs.
• Establish a benchmark dataset and baseline method for this new task.

Main Methods:

• Constructed a high-quality multi-view LF dataset (60 indoor, 59 outdoor scenes).
• Proposed a baseline method featuring adaptive alignment, multi-level feature decoupling, and refinement modules.
• Introduced an artifact-aware loss function for visual quality enhancement.

Main Results:

• The proposed method significantly outperforms existing approaches in sparse-to-dense inbetweening.
• Demonstrated enhanced LF view synthesis by filling interperspective gaps and increasing data robustness.
• Established a new benchmark for the sparse-to-dense inbetweening task.

Conclusions:

• The novel sparse-to-dense inbetweening task and baseline method advance LF imaging capabilities.
• The developed dataset and method provide a foundation for future research in LF view synthesis.
• The approach effectively handles sparse inputs and improves the robustness and quality of synthesized LF views.