Bridging Implicit and Explicit Geometric Transformation for Single-Image View Synthesis.

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• Single-image view synthesis aims to generate new viewpoints from a single input image.
• Autoregressive models achieve high-quality novel views but face a trade-off between content preservation and realistic completion (the "seesaw" problem) and are computationally expensive.

Purpose of the Study:

• To propose an efficient non-autoregressive framework for single-image view synthesis.
• To mitigate the "seesaw" problem by effectively combining explicit and implicit 3D geometry representations.
• To significantly reduce the computational cost associated with novel view generation.

Main Methods:

• Developed a novel loss function to complement explicit and implicit renderers.
• The loss function encourages explicit features to enhance reprojected areas and implicit features to improve out-of-view regions.
• Utilized an efficient non-autoregressive model architecture.

Main Results:

• Successfully alleviated the "seesaw" problem in single-image view synthesis.
• Achieved superior performance compared to autoregressive state-of-the-art methods.
• Generated novel views approximately 100 times faster than existing approaches.
• Validated effectiveness and efficiency on RealEstate10K and ACID datasets.

Conclusions:

• The proposed framework offers an effective and efficient solution for single-image view synthesis.
• The novel loss function successfully balances content preservation and realistic completion.
• Non-autoregressive models can achieve competitive or superior results with significantly reduced computational overhead.