Scalable Hole-Filling for Real-Time Multi-GPU Light Field Path Tracing.

Area of Science:

• Computer graphics
• Display technology
• Computational imaging

Background:

• Light field (LF) displays offer natural focus cues and motion parallax, overcoming limitations of traditional displays.
• LF path tracing is computationally intensive for real-time applications due to rendering multiple views.
• Pixel reprojection reduces complexity but creates rendering holes when occluded areas are uncovered, requiring additional computation.

Purpose of the Study:

• Investigate scalable hole-filling strategies for LF path tracing to achieve real-time performance.
• Optimize hole-filling algorithms using multi-GPU systems.
• Introduce and evaluate a novel hole-filling method, DaSH, for reduced computational overhead.

Main Methods:

• Developed an algorithm search optimization procedure to assess the generalizability of assignment algorithms for hole-filling.
• Introduced DaSH (Discarded and Subsampled path traced Hole-filling), a method that adaptively subsamples and discards rays based on local pixel sparsity.
• Utilized a multi-GPU rendering system for scalable performance evaluation.

Main Results:

• DaSH achieves significant performance gains in LF path tracing.
• The method preserves the visual and structural quality of refocused light field images.
• Demonstrated an average speedup factor of 1.8x compared to prior work in a multi-GPU system.

Conclusions:

• DaSH offers an effective and efficient solution for hole-filling in LF path tracing.
• The proposed method enables real-time performance for light field displays.
• Adaptive ray subsampling is a viable strategy for reducing computational complexity and divergence overhead.