ProbIBR: Fast Image-Based Rendering With Learned Probability-Guided Sampling.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Novel view synthesis is crucial for immersive experiences but often computationally expensive.
• Existing methods struggle with efficiency due to geometry pre-computation or exhaustive sampling in neural rendering.

Purpose of the Study:

• To develop a fast and practical solution for novel view synthesis from sparse input views.
• To improve the efficiency of neural volumetric rendering by reducing sampling points.

Main Methods:

• Incorporating learned Multi-View Stereo (MVS) priors into neural volume rendering.
• Employing probability-guided sampling based on depth distributions to reduce redundant points.
• Developing a confidence-aware refinement for uncertain and occluded regions.

Main Results:

• Achieved 15 to 40 times faster rendering compared to state-of-the-art methods.
• Demonstrated strong generalization capabilities across diverse real-world scenes.
• Rendered 512x512 novel views at ~20 fps on a single GTX 3090 GPU.

Conclusions:

• The proposed method offers a significant speedup for novel view synthesis.
• The approach maintains high-quality synthesis performance and generalizes well.
• Enables real-time free-viewpoint experiences and holographic display applications.